CN106250812A - A kind of model recognizing method based on quick R CNN deep neural network - Google Patents

Abstract

The present invention discloses a kind of model recognizing method based on quick R CNN deep neural network, mainly include the study of the unsupervised degree of depth, the CNN convolutional neural networks of multilamellar, region suggestion network, network share, softmax grader；Achieve a use truly quick R CNN real-time performance vehicle detection end to end and the framework of identification, and there is the Morphological Diversity being applicable to vehicle target, illumination variation multiformity, under the environment such as background multiformity quickly, high accuracy and the vehicle subclass identification of robustness.

Description

A kind of model recognizing method based on quick R-CNN deep neural network

Technical field

The present invention relates to computer technology, pattern recognition, artificial intelligence, applied mathematics and biological vision technology in intelligence The application of field of traffic, particularly relates to a kind of model recognizing method based on quick R-CNN deep neural network.

Background technology

Core Feature in intelligent transportation system is the accurately detection to vehicular traffic and correct vehicle cab recognition.The most right The research of vehicle detection sorting technique mainly has two important technologies: i.e. automatic vehicle identification and automobile automatic recognition.

Automatic vehicle identification is to utilize mobile unit to know mutually with ground base station equipment to carry out, and this technology is mainly used in charge system In system, relatively wide in some technology developed country ranges, such as AE-PASS system, the ETC system of Japan of the U.S., the whole world is defended Star GPS location etc..

Automobile automatic recognition is by detection vehicle parameter inherently, uses suitably under certain vehicle classification standard Classification and identification algorithm, on one's own initiative vehicle is carried out typing, this class technology Application comparison is extensive, and oneself through having a lot of ripe is System is applied in real life, and such technology can be known automatically by modes such as frequency microwave, HONGGUANG, laser, surface acoustic waves Other information of vehicles, it is possible to use the mode of Computer Vision identifies the information of vehicles such as car plate, vehicle.

Automobile automatic recognition comparative maturity technology has Data mining, swashs for infrared detection, ultrasound wave/microwave inspection Survey, geomagnetism detecting etc., but this several method is respectively arranged with quality, and advantage is that identification is the highest, but shortcoming is it is also obvious that mainly lack Point has construction and installation process sufficiently complex, affects normal traffic order, and difficult in maintenance, capital equipment is fragile, spends bigger Deng.

In recent years, video detection technology oneself become intelligent transportation field most important information gathering means, synthetical comparison and assessment, will Video detection technology is applied to highway and urban road has great actual application value, based on video vehicle cab recognition system System, by information gathering and the level of safety management of General Promotion urban road, can play increasingly in intelligent transportation system Important effect.

The visual identity of vehicle, lot of domestic and international scholar has carried out correlational study.Paper " Robert T.Collins, Alan J.Lipton,Hironobu Fujiyoshi,and Takeo Kanade.Algorith rns For Cooperative multisensor surveillanee.In Proceedings of the IEEE " disclose a road The detection of upper mobile target, follow the tracks of, identification system, identify that moving target is people, crowd, vehicle by the neutral net trained Or interference, the input characteristics amount of network has the dispersibility of target to measure, target sizes target surface size monitors with camera The relative value of area size.Dissimilar and color further divided into by vehicle.Paper " TieniuN.Tan and Keith D.Baker Efficient image gradient based vehicle localization.IEEE Transaction On Iimage Proeessing, 2000,9 (8): 1343-1356. " describe a kind of vehicle location and know method for distinguishing, one In individual wicket, the method is carried out according to image gradient.Utilize surface constraints and major part contour of the vehicle by two straight lines about The fact that bundle, the attitude of available vehicle.Paper " George Fung, NelsonYung, and Granthajm Pang.Vehicle shape approximation from motion for visual traffic Surveillance.In Proc.IEEE Conf.Intelligent Transport System, 2001,608-613. " use Vehicle shape is estimated in the motion of video camera observation vehicle in high precision, by estimating that characteristic point obtains vehicle's contour.Basic thought Being the translational speed translational speed more than low characteristic point of high characteristic point, because high characteristic point is close to video camera, vehicle's contour can With with vehicle identification.Paper " Ferryman, A.Worral, G.Sullivan, K.Baker, A generic deformable model for Vehicle recognition,Proeeeding of British Machine Vision Conference, 1995,127-136. " propose a parameterized deformable three-dimensional template, this template, by developing, it is said Can be applied to various vehicle identification.Paper " Larry Davis, V.Philomin and R.Duralswami.Tracking humans from a moving platform.In Proc.Intematl on Conference on Pattem Recognition, 2000. " study vehicle identification with deforming template, first, set up the side view of target vehicle vehicle head part And the deforming template of front view.By histogram intersection, the RGB rectangular histogram of vehicle also must compare, suitable vehicle template The point set on limit compares with other car modals also by the Hausdorff distance between point set.Above-mentioned technology the most also needs Manually to complete feature extraction, bigger problem is: 1) affected too big by concrete applied environment, all kinds of detection algorithm requirements Condition the harshest；2) vehicle class is various but difference is little, does not has obvious distinguishing characteristics；3) affected by visible change Greatly, the automobile characteristic difference taken the photograph from different perspectives is big；4) too big by natural environment influence, particularly illumination effect, serious Illumination reflection makes vehicle wheel profile fuzzy, and color deviation, change are the biggest, it is difficult to identification；5) profile of automobile updates too fast, Changing features is the fastest so that algorithm adaptability is poor.Domestic remain in research state at vehicle cab recognition technical elements great majority, Such as some achievements in research such as the Chinese Academy of Sciences, Xi'an Highway house, Shanghai Communications University, Xi'an Communications University, Sichuan Universitys.It closes Key problem is that the knowledge of vehicle cab recognition process is limited by the mankind itself.

It is used for carrying out the feature of vehicle classification to need yardstick, rotation and certain angular transformation, illumination variation are had Well robustness, computer vision technique typically all labor costs's plenty of time and the energy in front degree of depth study epoch go to set Count suitable feature.In order to be able to allow computer automatically select suitable characteristics, artificial neural network just arises at the historic moment, as far back as 1999 Just there is external researcher to utilize neutral net to sorting objects in year, include the side such as fuzzy neural network and BP network Method；But owing to its performance exists problems, as being difficult to solve complexity and the contradiction of generalization present in pattern recognition, god Although having powerful modeling ability through network, but classifying such as vehicle large-scale image, its huge parameter space makes Find excellent optimized initial value more difficulty etc., therefore treated coldly for a long time by people, until the proposition of degree of depth study is Become study hotspot again.

In image classification field, the image classification to substantial amounts mainly has two kinds of methods: a kind of for extracting every photo Local feature, carry out the feature of extraction clustering and coding obtain a high dimension vector, then it is classified with grader.Its The method of middle coding has visual word bag model to encode, and sparse coding and Fei Sheer vector coding etc., from the point of view of current result of study The performance of Fei Sheer vector coding to be got well compared with other several coded systems.The widest image classification method of another kind of application is the degree of depth Neutral net, degree of depth study is a new focus in neutral net research, its object is to by non-supervisory pre-training Excellent initial parameter value is provided for neutral net, by the way of greedy, training in layer, large-scale image is classified Obtain extraordinary effect.

It is before and after about 2006 that the concept of degree of depth study starts to attract much attention, paper " Hinton, G.E.and R.R.Salakhutdinov,Reducing the dimensionality of data with neural Networks.Science, 2006.313 (5786): 504-507 " feedforward neural network proposing a kind of multilamellar can be successively Do the training of efficient early stage, use unsupervised restricted Boltzmann machine that each layer is trained, finally have prison in utilization The back-propagating superintended and directed is finely tuned, and provides a kind of new side for solving the contradiction of complexity and generalization present in pattern recognition Method and thinking, thus pulled open the computer vision technique prelude in degree of depth study epoch.

Convolutional neural networks, i.e. CNN, be the one of degree of deep learning algorithm, is that the pattern in special disposal image domains is known Not, also it is the algorithm that in current image steganalysis, achievement is the most surprising simultaneously.Convolutional neural networks algorithm is advantageous in that training Need not the when of model use any manual features, algorithm can explore the feature that image is implied automatically.

The Chinese patent application of Application No. 201610019285.2 discloses a kind of model recognizing method and system, including Machine training generates grader process and treats the differentiation process of mapping sheet, during generating grader, based on car plate Determine in training set picture required image range, by it has been determined that image range zoning, in selected each region All characteristic informations in selected respective region are put into machine training and generate dividing of each region of one_to_one corresponding by characteristic information respectively Class device, treats mapping sheet by the grader generated and carries out single area judging, according to single area judging result again through too much Region confidence merges judgement and obtains vehicle cab recognition result.This invention uses random deep woods grader to carry out vehicle cab recognition, maximum Problem be the absence of unsupervised learning process.

The Chinese patent application of Application No. 201510639752.7 discloses a kind of model recognizing method, described method bag Include: obtain picture to be detected；Use first to preset grader described picture to be detected is detected；If described picture to be detected In containing target vehicle, extract the target vehicle in described picture to be detected；Described target vehicle is carried out registration process, so that Angle between headstock direction and the vertical direction of described target area of described target vehicle is less than predetermined threshold value；To described right Target vehicle after neat process carries out feature extraction, and to obtain M feature, described M is the integer more than 1；Use second pre- If described M feature is classified by grader；Result according to described classification determines the vehicle of described target vehicle.From certain Saying in meaning, this invention still falls within shallow neutral net, is the most just difficult to unsupervised learning process.

The Chinese patent application of Application No. 201510071919.4 proposes a kind of vehicle based on convolutional neural networks Recognition methods, feature based extraction module and vehicle cab recognition module, comprise the following steps: by design convolution and pond layer, entirely Articulamentum, grader builds the neutral net of vehicle cab recognition, and wherein convolution is used for extracting vehicle with pond layer and full articulamentum Feature, grader is used for vehicle classification identification；Utilize this neutral net of database training comprising different automobile types feature, training side Formula is the study having supervision that the data of tape label are carried out, and carries out weight parameter matrix and side-play amount by stochastic gradient descent method Adjustment；Obtain the weight parameter matrix in each layer trained and side-play amount, they are assigned to accordingly this neutral net In each layer, then this network has the function of vehicle feature extraction and identification.This invention lacks the ratio details that realizes in greater detail, Simply conceptually proposing model recognizing method based on convolutional neural networks, especially training method is the data of tape label The study having supervision carried out.The profile image volume of vehicle belongs to mass data, and will be labeled these view data is one Individual extremely difficult thing；In addition the profile of vehicle updates and causes changing features fast, so causing some in this invention the soonest Algorithm is difficult to meet reality application.Additionally, the vehicle image that actual road surface collection arrives is complicated；Include background complicated, vehicle it Between the interference such as block；If the image without obtaining reality carries out dividing processing, it will have a strong impact on final recognition result.

The Chinese patent application of Application No. 201510738852.5 and 201510738843.6 proposes a kind of based on deeply The model recognizing method of Du Feisheer network, first builds the 0th layer of Fei Sheer network, to there being the data base of K kind vehicle image, Extract the SIFT feature of every kind of vehicle vehicle image；Then the 1st layer of Fei Sheer network is built, to the SIFT feature extracted Carry out Fei Sheer vector coding, the vector after coding is stacked in space, then carries out L2 normalization and PCA dimensionality reduction；1st layer is obtained To character representation carry out Fei Sheer vector coding, by symbol square root and L2 normalized, form Fei Sheer network 2nd layer；Finally the global characteristics that different automobile types image obtains is represented and be used for linear SVM training, obtain that there is K kind The identification system of vehicle classification；To vehicle to be identified so that it is obtain testing feature vector by Fei Sheer network, import and identify system System i.e. may recognize that vehicle vehicle to be identified.This invention also has two deficiencies, and the first lacks degree of deep learning process, and it two is scarce Few vehicles segmentation that carries out image processes step.

The Chinese patent application of Application No. 201510738540.4 proposes a kind of based on local feature Aggregation Descriptor Model recognizing method, first extract the SIFT feature of vehicle image in model data storehouse；Then to all vehicle images SIFT feature carries out Kmeans cluster, forms K cluster centre, obtains the dictionary with K vision word；Then for every Each SIFT feature is assigned to the vision word of its nearest neighbours by vehicle image；Add up SIFT feature around each vision word to Amount and the residual error cumulant of Current vision word, obtain the local feature Aggregation Descriptor of Current vehicle picture；Finally, will training The n of module opens the local feature Aggregation Descriptor of vehicle image, by quantization encoding, obtains a n class vehicle class that can index Other coded image storehouse；And to test vehicle image, same its local feature Aggregation Descriptor that extracts, as query vector, lead Enter image library to be indexed, mated by approximate KNN searching method, identify test vehicle vehicle.This invention same Also having two defects, the first lacks degree of depth study, and it two is the absence of carrying out image vehicles segmentation and processes step.

Convolutional neural networks is in truck, buggy, big bus, minibus, SUV and car identification or compares success , but classify in subclass, the precision less than the classification of big class as the most remote in the precision in the different automobile types of identification car.As a rule The difficulty of subclass image recognition is generally placed at 2 points:

(1) acquisition of subclass image labeling data is the most difficult, it usually needs the expert of association area is labeled.

(2) difference in subclass image recognition exists big class, as in vehicle cab recognition, the viewing angle that car is different, and Difference between little class.

In sum, use convolutional neural networks even depth nerual network technique to vehicle cab recognition, the most still also exist as Several stubborn problems lower: 1) from complicated background, how to be accurately partitioned into the general image of tested vehicle；2) the most to the greatest extent Few label image data may be used accurately to obtain the characteristic of vehicle vehicle；3) how to know in the big class of vehicle vehicle Also can recognize which kind of is on the basis of Bie, car which kind of color, which go out in age；4) how to learn automatically to obtain by the degree of depth The feature of a vehicle of picking up the car；5) how to take into account accuracy of identification and detection efficiency, reduce training and learning time the most as far as possible； 6) how when classifier design, this grader can meet the classificating requirement of vehicle vehicle subclass, again can be in the profile of automobile Without the most whole network being trained study after renewal；7) one CNN of a use truly how is designed Real-time performance vehicle detection end to end and the framework of identification；8) how to reduce the impact of weather condition, increase the adaptive of system Ying Xing.

Summary of the invention

In order to overcome the automatization in existing vehicle vehicle Visual identification technology and intelligent level is low, lack the degree of depth Practise, be difficult to adapting to ambient weather change, being difficult to accurately to extract the vehicle general image for identifying, be difficult to use visual manner Vehicle vehicle subclass being identified classification, is difficult to take into account accuracy of identification and the deficiency such as time and detection efficiency, the present invention provides A kind of model recognizing method based on quick R-CNN deep neural network, can be effectively improved vehicle visual identity automatization and Intelligent level, can preferably adapt to ambient weather change have widely adaptivity, can guarantee that and identify essence in preferably detection Have on the basis of degree and detect identification ability in real time, the dependence to label vehicle data can be greatly reduced there is automatically study and extract car The ability of type feature, the contradiction of the complexity and generalization that can preferably solve vehicle cab recognition have preferably universality.

Foregoing invention content to be realized, it is necessary to solve several key problem: (1) designs quickly regarding of a kind of Vehicle Object Feel partitioning algorithm；(2) a kind of degree of deep learning method of research and development, it is achieved unsupervised vehicle feature extraction；(3) design one is applicable to The grader of the type subclass of thousands of kinds, and there is autgmentability；(4) design one quick R-of one use truly CNN real-time performance vehicle detection end to end and the framework of identification.

The technical solution adopted for the present invention to solve the technical problems is:

A kind of model recognizing method based on quick R-CNN deep neural network, including one for degree of depth study and instruction Practice the VGG network identified, a region suggestion network being used for extracting area-of-interest and one for vehicle classification Softmax grader；

Described VGG network, including 8 Ge Juan basic units, 3 full articulamentums, 11 layers altogether；8 Ge Juan basic units there are 5 organize Convolutional layer, 2 classification layers extract characteristics of image, 1 classification layer characteristic of division；3 full articulamentums are distinguished link sort layer 6, are divided Class layer 7 and classification layer 8；

Described region suggestion network, returns layer, 1 module calculating Classification Loss including 1 classification layer, 1 window With the module that 1 calculation window returns loss, p Suggestion box interested of output；

Described Softmax grader, obtains feature database data by the input data characteristics and the learning training that extract and enters Row comparison, calculates the probability of each classification results, and the result then taking probability the highest exports；

Quickly R-CNN deep neural network, has accessed described region suggestion at the 5th layer of end of described VGG network Network so that described region suggestion network shares low-level image feature extraction process and the result of first 5 layers of described VGG network；

The the 6th and the 7th layer of described VGG network is according to p suggestion interested of described region suggestion network output Characteristics of image in frame carries out convolution and ReLU process, obtains p the characteristic pattern containing 4096 vectors, gives classification the most respectively Layer and window return layer and process, it is achieved the segmentation of vehicle image；On the other hand, p is contained by described Softmax grader The characteristic pattern having 4096 vectors carries out Classification and Identification, obtains the classification results of vehicle vehicle.

Described Softmax grader, using the learning outcome in quick R-CNN as softmax during learning training The input data of grader；It is that the Logistic towards multicategory classification problem returns that Softmax returns, and is that Logistic returns General type, it is adaptable to the situation of mutual exclusion between classification；Assume for training set { (x⁽¹⁾,y⁽¹⁾,…,x^(m),y^(m)), there is y⁽¹⁾ ∈ 1,2 ..., k}, for given sample input x, the vector exporting a k dimension represents what each classification results occurred Probability is p (y=i | x), it is assumed that function h (x) is as follows:

$h_{\mathrm{\θ}}\left(x^{\left(i\right)}\right)=\left[\begin{array}{c}p(y^{\left(i\right)}=1|x^{\left(i\right)},\mathrm{\θ})\\ p(y^{\left(i\right)}=1|x^{\left(i\right)},\mathrm{\θ})\\ .\\ .\\ .\\ p(y^{\left(i\right)}=k|x^{\left(i\right)},\mathrm{\θ})\end{array}\right]=\frac{1}{{\Σ}_{j=1}^k{e}^{{\mathrm{\θ}}_j^T{x}^{\left(i\right)}}}\left[\begin{array}{c}{e}^{{\mathrm{\θ}}_1^T{x}^{\left(i\right)}}\\ e^{{\mathrm{\θ}}_2^T{x}^{\left(i\right)}}\\ .\\ .\\ .\\ e^{{\mathrm{\θ}}_k^T{x}^{\left(i\right)}}\end{array}\right]---\left(11\right)$

θ₁,θ₂,…θ_kIt is the parameter of model, and all of probability and be 1；Adding the cost function after regularization term is:

$J\left(\mathrm{\θ}\right)=-\frac{1}{m}\[\underset{i=1}{\overset{m}{\Σ}}\underset{j=1}{\overset{k}{\Σ}}1\{y^{\left(i\right)}=j\}\log \frac{e^{{\mathrm{\θ}}_j^T{x}^{\left(i\right)}}}{{\Σ}_{l=1}^k{e}^{{\mathrm{\θ}}_l^T{x}^{\left(i\right)}}}\]+\frac{\mathrm{\λ}}{2}\underset{l=1}{\overset{k}{\Σ}}\underset{j=0}{\overset{n}{\Σ}}{\mathrm{\θ}}_{ij}^2---\left(12\right)$

The partial derivative of l parameter of jth classification is by cost function:

${\▿}_{{\mathrm{\θ}}_j}J\left(\mathrm{\θ}\right)=-\frac{1}{m}\underset{i=1}{\overset{m}{\Σ}}\[x^{\left(i\right)}\left(1\right\{y^{\left(i\right)}=j\}-p(y^{\left(i\right)}=j|x^{\left(i\right)};\mathrm{\θ}\left)\right)\}\]+{\mathrm{\λ\θ}}_j---\left(13\right)$

Finally, by minimizing J (θ), it is achieved the classification of softmax returns, and classification regression result is saved in feature database In；

When identifying classification, the input data characteristics and the learning training that extract are obtained feature database data and compare, Calculating the probability of each classification results, the result then taking probability the highest exports.

Described region suggestion network, for formation zone Suggestion box, accesses at the 5th layer of end of described VGG network Region suggestion network described in, the convolution Feature Mapping figure that i.e. convolutional layer at the 5th layer exports slides a little network, this Individual network is connected to input in the spatial window of the n × n of convolution Feature Mapping entirely；Each sliding window be mapped to a low-dimensional to In amount, low dimensional vector is 256-d, the corresponding numerical value of a sliding window of each Feature Mapping；This vector output is to two The full layer connected of individual peer；Individual window returns layer and a classification layer；Window returns layer and exports on each position, 9 kinds Recommending region correspondence window to need have translation scaling invariance, window returns layer and exports 4 translation scalings from 256 dimensional features Parameter, has 4k output, the i.e. codes co-ordinates of k Suggestion box；Classification layer exports from 256 dimensional features and belongs to foreground and background Probability, exports 2k Suggestion box score, is i.e. the estimated probability of vehicle target/non-vehicle target to each Suggestion box.

The most whether the training of region suggestion network, distribute a binary label to each candidate region, be vehicle pair As；Here positive label is distributed to two class candidate regions: (i) and the enclosing region of certain GT have the ratio of the highest common factor union, IoU, overlapping candidate region；(ii) candidate region overlapped with any GT enclosing region IoU more than 0.7；Distribution is negative simultaneously Label is below the candidate region of 0.3 to the IoU ratio with all GT enclosing region；Leave out the candidate region of anon-normal non-negative；Tool Body algorithm is as follows:

STEP31: order reads every figure in training set；

STEP32: the true value candidate region to each demarcation, the candidate region overlapping ratio maximum is designated as prospect sample This；

STEP33:: to STEP32) remaining candidate region, if it is overlapping with certain demarcation, IoU ratio is more than 0.7, It is designated as prospect sample；If its overlap proportion demarcated with any one is both less than 0.3, it is designated as background sample；

STEP34: candidate region remaining to STEP32 and STEP33 is discarded；

STEP35: the candidate region crossing over image boundary is discarded.

In order to automatically carry out screening and the regional location refine of candidate region, use here and minimize object function；To one The cost function formula (14) of individual image represents,

$L\left(\right\{p_i\},\{t_i\left\}\right)=\frac{1}{N_{cls}}\underset{i}{\Σ}{L}_{cls}(p_i,p_i^{*})+\mathrm{\λ}\frac{1}{N_{reg}}\underset{i}{\Σ}{p}_i^{*}{L}_{reg}(t_i,t_i^{*})---\left(14\right)$

In formula, i is the index of candidate region, N in a batch processing_clsFor the normalization coefficient of layer of classifying, N_regFor window Returning the normalization coefficient of layer, λ is balance weight, p_iFor the prediction probability of vehicle target,For GT label, if candidate region For justIf candidate region is negativet_iIt is a vector, represents 4 the parametrization coordinates surrounding frame of prediction,The coordinate vector of frame, L is surrounded for the GT corresponding with positive candidate region_clsFor the logarithm cost of classification, L_regFor returning logarithm generation Valency, L ({ p_i},{t_i) it is total logarithm cost；

Logarithm cost L of classification_clsCalculated by formula (15),

$L_{cls}(p_i,p_i^{*})=-log\[p_i^{*}{p}_i+(1-p_i^{*})(1-p_i)\]---\left(15\right)$

Window returns logarithm cost L_regCalculated by formula (16),

$L_{reg}(t_i,t_i^{*})=R(t_i-t_i^{*})---\left(16\right)$

In formula, R is the cost function of the robust of definition, belongs to Smooth L1 error, insensitive to outlier, uses formula (17) calculate,

In formula (14)This means only positive candidate region, i.e.Shi Caiyou returns generation Valency, other situations due toDo not return cost；Classification layer and window return the output of layer respectively by { p_iAnd { t_iGroup Becoming, these two respectively by N_clsAnd N_regAnd a balance weight λ normalization, select λ=10, N here_cls=256, N_reg= 2400, returning layer item by such selection sort layer and window is almost equal weight；

About position refine, using 4 values, centre coordinate, width and height here, computational methods are as follows,

$\begin{array}{c}{t}_x=(x-x_a)/w_a,t_y=(y-y_a)/h_a,t_w=\log (w/w_a),t_h=\log (h/h_a),\\ {t^{*}}_x=(x^{*}-x_a)/w_a,{t^{*}}_y=(y^{*}-y_a)/h_a,{t^{*}}_w=\log (w^{*}/w_a),{t^{*}}_h=\log (h^{*}/h_a)\end{array}---\left(18\right)$

In formula, x, y, w, h represent encirclement frame centre coordinate, width and height, x respectively_a、y_a、w_a、h_aRepresent candidate respectively Regional center coordinate, width and height, x^*、y^*、w^*、h^*Represent the encirclement frame centre coordinate of prediction, width and height respectively；With The result of calculation of formula (18) carries out position refine；It practice, explicitly do not extract any candidate window, use district completely Territory suggestion network self completes to judge and position refine.

Described VGG network, the method setting up multilayer neural network in label vehicle image data, it is divided into two steps, one Being every time training one layer network, two is tuning, and senior expression r and this senior expression r making original representation X upwards generate gives birth to downwards The X' become is the most consistent；

The propagated forward process of convolutional neural networks, the output of last layer is the input of current layer, and by activating letter Number successively transmits, and Practical Calculation output formula (4) of the most whole network represents,

O_p=F_n(…(F₂(F₁(XW₁)W₂)…)W_n) (4)

In formula, X represents and is originally inputted, F_lRepresent the activation primitive of l layer, W_lRepresent the mapping weight matrix of l layer, O_p Represent the Practical Calculation output of whole network；

The output of current layer represents with (5),

X^l=f^l(W^lX^l-1+b^l) (5)

In formula, l represents the network number of plies, X^lRepresent the output of current layer, X^l-1Represent the output of last layer, i.e. current layer Input, W^lRepresent trained, the mapping weight matrix of current network layer, b^lAdditivity for current network is bigoted, f^lIt is to work as The activation primitive of front Internet；The activation primitive f used^lFor correcting linear unit, i.e. ReLU, represent with formula (6),

$f^l=max({\left(W^l\right)}^T{X}^l,0)=\left\{\begin{array}{cc}{\left(W^l\right)}^T{X}^l& {\left(W^l\right)}^T{X}^l>0\\ 0& {\left(W^l\right)}^T{X}^l\≤0\end{array}\right.---\left(6\right)$

In formula, l represents the network number of plies, W^lRepresent trained, the mapping weight matrix of current network layer, f^lIt is to work as The activation primitive of front Internet；Its effect is that then allowing it is 0 if convolutional calculation result is less than 0；Otherwise keep its value constant.

Described VGG network, first 5 layers is a typical degree of depth convolutional neural networks, and this neural metwork training is one Back-propagation process, by error function back propagation, utilizes stochastic gradient descent method to be optimized deconvolution parameter and biasing Adjust, until network convergence or reach maximum iteration time stop；

Back propagation needs by comparing the training sample with label, uses square error cost function, right In c classification, the multi-class of N number of training sample is identified, and network final output error function formula (7) calculates by mistake Difference,

$E^N=\frac{1}{2}\underset{n=1}{\overset{N}{\Σ}}\underset{k=1}{\overset{c}{\Σ}}{(t_k^n-y_k^n)}^2---\left(7\right)$

In formula, E^NFor square error cost function,It is the kth dimension of the n-th sample corresponding label,It it is the n-th sample The kth output of map network prediction；

When error function is carried out back propagation, use computational methods as traditional BP class of algorithms, such as formula (8) institute Show,

$\begin{array}{c}{\mathrm{\δ}}^l={\left(W^{l+1}\right)}^T{\mathrm{\δ}}^{l+1}\×f^{\′}\left(u^l\right)\\ u^l=W^l{x}^{l-1}+b^l\end{array}---\left(8\right)$

In formula, δ^lRepresent the error function of current layer, δ^l+1Represent the error function of last layer, W^l+1Square is mapped for last layer Battle array, f' represents the inverse function of activation primitive, i.e. up-samples, u^lRepresent the output not by the last layer of activation primitive, x^l-1Represent The input of next layer, W^lWeight matrix is mapped for this layer；

After error back propagation, obtain the error function δ of each Internet^l, then use stochastic gradient descent method To network weight W^lModify, then carry out next iteration, until network reaches the condition of convergence；Need when carrying out error propagation Up-sampling in formula to be first passed through (8) makes before and after's two-layer equivalently-sized, then carries out error propagation；

Algorithm idea is: 1) the most successively build monolayer neuronal unit, is one single layer network of training the most every time；2) when After all layers have been trained, wake-sleep algorithm is used to carry out tuning.

Degree of deep learning training process is specific as follows:

STEP21: use unsupervised learning from bottom to top, i.e. from the beginning of bottom, past top layer training in layer, learn Practise vehicle image feature: first train ground floor with without label vehicle image data, during training, first learn the parameter of ground floor, due to Model holds quantitative limitation and sparsity constraints so that the model obtained can learn the structure to data itself, thus obtains The feature of expression ability is had more than input；After study obtains l-1 layer, using the output of l-1 layer as the input of l layer, Train l layer, thus respectively obtain the parameter of each layer；Specifically calculate as shown in formula (5), (6)；

STEP22: top-down supervised learning, i.e. by the vehicle image data of tape label go training, error from top to Lower transmission, is finely adjusted network: specifically calculate as shown in formula (7), (8)；

The each layer parameter obtained based on STEP21 finely tunes the parameter of whole multilayered model further, and this step is one prison Supervise and instruct experienced process；STEP21 is similar to the random initializtion initial value process of neutral net, due to the degree of depth study STEP21 be not with Machine initializes, but obtained by the structure of study input data, thus this initial value is closer to global optimum such that it is able to Obtain more preferable effect.

The model initialization of first 5 layers of described VGG network: be broadly divided into data and prepare, calculate image average, network Define, train and recover 5 steps such as data；

1) data prepare；Be have collected the view data of all kinds of vehicle by reptile software, obtain is substantially with mark The vehicle image data signed, as training image data；Another kind of data are the vehicle figures obtained by bayonet camera As data；

2) image average is calculated；Model needs to deduct average from every pictures；

3) definition of network；Main definitions xml tag path, the path of picture, deposit train.txt, val.txt, The path of test.txt and trainval.txt file；

4) training；Run training module；

5) data are recovered；The layer of ReLu5 before deleting, and change the bottom of roi_pool5 into data and rois；

The model initialization work of vehicle vehicle pre-training is completed through above-mentioned process.

Described region advises that front 5 layers of low-level image feature of the VGG network described in network utilisation extract result, i.e. two nets Network have shared front 5 layers of low-level image feature of described VGG network, needs to learn to optimize the feature shared by alternative optimization；Tool Body algorithm is as follows:

STEP41: with the optimization training region suggestion network of region suggestion network, prepare by above-mentioned data, calculate image Average, the definition of network, train and recover 5 steps such as data and complete model initialization, and end-to-end fine setting is built for region View task；

STEP42: the Suggestion box generated with the region suggestion network of STEP1, is trained a single inspection by quick R-CNN Survey grid network, this detection network is by the model initialization of vehicle vehicle pre-training equally, and at this time two networks also do not have Shared volume lamination；

STEP43: with detection netinit region suggestion network training, fix the convolutional layer shared, and only finely tune district The layer that territory suggestion network is exclusive, at this moment two network shared volume laminations；

STEP44: keep the convolutional layer shared to fix, finely tune the classification layer of quick R-CNN；So, two networks share phase Same convolutional layer, finally constitutes a unified network.

In the present invention, vehicle vehicle visual identity main flow is as follows；

STEP51: read image to be identified；

STEP52: be normalized image to be identified, obtains tri-different colours 224 × 224 of RGB normalized View data；

STEP53: tri-normalized view data of different colours 224 × 224 of RGB are input to three CNN passages, through 5 Layer process of convolution obtains vehicle vehicle character image data；

STEP54: the Suggestion box generated vehicle vehicle character image data by region suggestion network, chooses one The Suggestion box of high score, i.e. obtains an area-of-interest, RoI；By maximum 5 layers of pondization, this RoI is carried out process obtain The trellis diagram of one 6 × 6 × 256RoI；

STEP55: the trellis diagram of RoI is exported the feature obtaining 4096 dimensions to two full layers connected at the same level after processing Vector, as the input data of softmax grader；

STEP56: the classification regression analysis to characteristic vector softmax of 4096 dimensions obtains vehicle vehicle cab recognition result, Identify by which kind of vehicle the vehicle in altimetric image belongs to.

Beneficial effects of the present invention is mainly manifested in:

1) a kind of model recognizing method based on quick R-CNN deep neural network is provided；

2) a kind of degree of deep learning method of research and development, it is achieved unsupervised vehicle feature extraction；

3) design the grader of a kind of type subclass being applicable to thousands of kinds, and there is autgmentability；

4) the quick R-CNN real-time performance vehicle detection end to end of a use truly and knowledge are achieved Other framework, and there is the Morphological Diversity being applicable to vehicle target, illumination variation multiformity, fast under the environment such as background multiformity Speed, high accuracy and the vehicle subclass identification of robustness.

Accompanying drawing explanation

Fig. 1 is the detection algorithm flow process of marginal information candidate frame；

Fig. 2 is the process content in region suggestion network；

Fig. 3 is the sliding window schematic diagram of 3 kinds of yardsticks and 3 kinds of length-width ratios；

Fig. 4 is the synoptic diagram of region suggestion network；

Fig. 5 is that region Suggestion box generates explanatory diagram；

Fig. 6 is the explanatory diagram of the shared network in quick R-CNN network；

Fig. 7 is by the candidate region obtained after region suggestion network processes to vehicle on real road；

Fig. 8 is for completing to judge and position refine schematic diagram with cost function feasible region suggestion network self；

Fig. 9 is vehicle cab recognition FB(flow block) based on CNN；

Figure 10 is wake-sleep algorithmic descriptions figure；

Figure 11 is the vehicle cab recognition FB(flow block) of the CNN of VGG model；

Figure 12 is quick R-CNN real-time performance vehicle detection end to end and the training procedure chart of identification；

Figure 13 is quick CNN real-time performance vehicle detection end to end and identification process synoptic diagram；

Figure 14 is quick R-CNN real-time performance vehicle detection end to end and the flow chart of identification.

Detailed description of the invention

Below in conjunction with the accompanying drawings technical scheme is described in further detail.

Embodiment 1

With reference to Fig. 1～14, the technical solution adopted for the present invention to solve the technical problems is:

(1) about the fast vision partitioning algorithm designing a kind of Vehicle Object；

First, design the fast vision partitioning algorithm of a kind of Vehicle Object, i.e. Vehicle Object is carried out regional choice and determines Position；

In order to the position of vehicle target is positioned；Owing to vehicle target possibly be present at any position of image, and And the size of target, Aspect Ratio are the most uncertain, original technology is that entire image is entered by the strategy of original adoption sliding window Row traversal, and need to arrange different yardsticks, different length-width ratios；Although this exhaustive strategy contains that target is all can The position that can occur, but shortcoming is also apparent from: and time complexity is the highest, produces redundancy window too many, and this is the most serious Affect subsequent characteristics extraction and the speed of classification and performance；

The problem existed for sliding window, the present invention proposes the solution of a kind of candidate region；Find out the most in advance The position that in figure, vehicle target is likely to occur；The information such as the texture in image, edge, color, energy is make use of due to candidate region Ensure to keep higher recall rate in the case of choosing less window；The time that so can effectively reduce subsequent operation is complicated Degree, and the candidate window obtained is higher than the quality of sliding window；Available algorithm is selective search, i.e. Selective Search and marginal information candidate frame, i.e. edge Boxes；The core of these algorithms is to make use of human vision " take a panoramic view of the situation " at a glance, directly find the vehicle target " general position " in entire image；Owing to selective search is calculated Method is the biggest, inapplicable Yu online vehicle cab recognition and detection；The present invention uses the detection algorithm of marginal information candidate frame.

The detection algorithm thought of marginal information candidate frame is: utilize marginal information, determine profile number in candidate frame and With the profile number of candidate frame imbricate, and based on this, candidate frame is marked, further according to the sequence of score Determine by size, length-width ratio, the candidate region information that position is constituted；Detection algorithm flow process such as Fig. 1 institute of marginal information candidate frame Show；Algorithm steps is as follows:

STEP11: processing original image with structure deep woods edge detection algorithm, the edge image obtained, then with non- The process further to edge image of maximum Restrainable algorithms obtains a edge image the most sparse；

STEP12: by being close to marginal point point-blank in the most sparse edge image, put together formation one Individual edge group, concrete way is, ceaselessly finds the marginal point of 8 connections, poor until the orientation angle between marginal point two-by-two Value and more than pi/2, the most just obtained many edge groups s of N_i∈S；

STEP13: calculate the similarity between two two edges groups with formula (1),

a(s_i,s_j)=| cos (θ_i-θ_ij)cos(θ_j-θ_ij)|^γ(1)

In formula, θ_iAnd θ_jIt is respectively the average orientation of two edge groups, s_iAnd s_jRepresent two edge groups, θ respectively_ijIt is two Mean place x of individual edge group_iAnd x_jBetween angle, γ is similar sensitivity coefficient, selects γ=2, a (s here_i,s_j) two Similarity between edge group；In order to improve computational efficiency, here by similarity a (s_i,s_j) value of calculation exceedes threshold value, T_s≥ The edge group of 0.05 stores, and remaining is disposed as zero；

STEP14: giving weights to each edge group, weight calculation method is given by formula (2),

$W_b\left(s_i\right)=1-\underset{T}{max}\underset{j}{\overset{\left|T\right|-1}{\Π}}a(t_j,t_{j+1})---\left(2\right)$

In formula, T is that the edge from candidate frame starts to arrive s_iThe path of edge group arrangement set, W_b(s_i) it is edge s_i Weights, t_jFor ..., (parameter interpretation)；Without finding path just by W_b(s_i) it is set as 1；

STEP15: calculate the scoring of candidate frame with formula (3),

$h_b=\frac{{\Σ}_i{W}_b\left(s_i\right)m_i}{2{(b_w+b_h)}^k}---\left(3\right)$

In formula, m_iFor in edge group s_iIn size m of all edge p_pSummation, W_b(s_i) it is edge s_iWeights, b_w And b_hIt is respectively width and height, the coefficient sized by k of candidate frame, defines k=1.5 here；Calculation window inward flange number is entered Row marking, last Ordering and marking filters out the candidate frame of low point；For present invention is mainly applied to the extraction of bayonet vehicle, this In just select the candidate frame of best result as tested vehicle object foreground image；

(2) about a kind of degree of deep learning method of research and development, it is achieved unsupervised vehicle feature extraction；

Due to the Morphological Diversity of vehicle target, illumination variation multiformity, the factor such as background multiformity makes to design one The feature of robust is not the easiest；But the quality extracting feature directly influences the accuracy of classification；

Meeting above three diversity requirement, robust vehicle feature extraction must be by coming real without supervision degree of depth study Now completing, successively initializing is a highly useful solution；The degree of depth study essence, be had by structure the most hidden The machine learning model of layer and the training data of magnanimity, learn more useful feature, thus finally promotes classification or prediction Accuracy；Therefore, the present invention realizing the degree of depth by following 2 to learn: 1) degree of depth of model structure is of five storeys～10 multilamellars Hidden node；2) by successively eigentransformation, the sample character representation in former space is transformed to a new feature space, from And make classification or prediction be more prone to；

The present invention, it is proposed that a kind of in non-supervisory data, i.e. without setting up multilamellar nerve net in label vehicle image data The method of network, briefly, is divided into two steps, and one is every time training one layer network, and two is tuning, makes original representation X upwards generate Senior expression r the most consistent with the X' that this senior expression r generates downwards；

The propagated forward process of convolutional neural networks, the output of last layer is the input of current layer, and by activating letter Number successively transmits, and Practical Calculation output formula (4) of the most whole network represents,

O_p=F_n(…(F₂(F₁(XW₁)W₂)…)W_n) (4)

In formula, X represents and is originally inputted, F_lRepresent the activation primitive of l layer, W_lRepresent the mapping weight matrix of l layer, l= 1,2,3 ... represent the network number of plies, O_pRepresent the Practical Calculation output of whole network；

The output of current layer formula (5) represents,

X^l=f^l(W^lX^l-1+b^l) (5)

In formula, l represents the network number of plies, X^lRepresent the output of current layer, X^l-1Represent the output of last layer, i.e. current layer Input, W^lRepresent trained, the mapping weight matrix of current network layer, b^lAdditivity for current network is bigoted, f^lIt is to work as The activation primitive of front Internet；The activation primitive f that the present invention uses^lFor correcting linear unit, i.e. Rectified Linear Units, ReLU, represent with formula (6),

$f^l=max({\left(W^l\right)}^T{X}^l,0)=\left\{\begin{array}{cc}{\left(W^l\right)}^T{X}^l& {\left(W^l\right)}^T{X}^l>0\\ 0& {\left(W^l\right)}^T{X}^l\≤0\end{array}\right.---\left(6\right)$

In formula, l represents the network number of plies, W^lRepresent trained, the mapping weight matrix of current network layer, f^lIt is to work as The activation primitive of front Internet；Its effect is that then allowing it is 0 if convolutional calculation result is less than 0；Otherwise keep its value constant；

The method using local receptor field and weights to share in CNN reduces network parameter further.So-called local Receptive field, refers to every kind of convolution kernel and is only connected with certain specific region in image, i.e. every kind convolution kernel convolved image A part, the most again in other layers by these local convolution features link together, both met image pixel in space On relatedness, reduce again deconvolution parameter quantity.Weights are shared and are then so that the weights of every kind of convolution kernel are the most identical, pass through The kind increasing convolution kernel extracts image many-side feature；In order to provide more details special to the classification of vehicle vehicle subclass Levy, the present invention appropriately increases the kind of convolution kernel；

Convolutional neural networks training is a back-propagation process, similar with traditional BP algorithm, anti-by error function To propagation, utilize stochastic gradient descent method that deconvolution parameter and biasing are optimized and revised, until network convergence or reach Big iterations stops.

Back propagation needs by comparing the training sample with label, calculates error.For example with a square mistake Difference cost function, for c classification, the multi-class identification problem of N number of training sample, network final output error function formula (7) represent,

$E^N=\frac{1}{2}\underset{n=1}{\overset{N}{\Σ}}\underset{k=1}{\overset{c}{\Σ}}{(t_k^n-y_k^n)}^2---\left(7\right)$

In formula, E^NFor square error cost function,It is the kth dimension of the n-th sample corresponding label,It it is the n-th sample pair The kth answering neural network forecast exports；

When error function is carried out back propagation, use computational methods as traditional BP class of algorithms, such as formula (8) institute Show,

$\begin{array}{c}{\mathrm{\δ}}^l={\left(W^{l+1}\right)}^T{\mathrm{\δ}}^{l+1}\×f^{\′}\left(u^l\right)\\ u^l=W^l{x}^{l-1}+b^l\end{array}---\left(8\right)$

In formula, δ^lRepresent the error function of current layer, δ^l+1Represent the error function of last layer, W^l+1Square is mapped for last layer Battle array, f' represents the inverse function of activation primitive, i.e. up-samples, u^lRepresent the output not by the last layer of activation primitive, x^l-1Represent The input of next layer, W^lWeight matrix is mapped for this layer；

After error back propagation, obtain the error function δ of each Internet^l, then use stochastic gradient descent method To network weight W^lModify, then carry out next iteration, until network reaches the condition of convergence；It should be noted that due to layer Different from the size dimension between layer, need the up-sampling first passed through in formula (8) to make before and after two when carrying out error propagation Layer is equivalently-sized, then carries out error propagation；

Algorithm idea is: 1) the most successively build monolayer neuronal unit, is one single layer network of training the most every time；2) when After all layers have been trained, wake-sleep algorithm is used to carry out tuning.

Degree of deep learning training process is specific as follows:

STEP21: use unsupervised learning from bottom to top, i.e. from the beginning of bottom, past top layer training in layer, learn Practise vehicle image feature: first train ground floor with without label vehicle image data, during training, first learn the parameter of ground floor, due to Model holds quantitative limitation and sparsity constraints so that the model obtained can learn the structure to data itself, thus obtains The feature of expression ability is had more than input；After study obtains l-1 layer, using the output of l-1 layer as the input of l layer, Train l layer, thus respectively obtain the parameter of each layer；Specifically calculate as shown in formula (5), (6)；

STEP22: top-down supervised learning, i.e. by the vehicle image data of tape label go training, error from top to Lower transmission, is finely adjusted network: specifically calculate as shown in formula (7), (8)；

The each layer parameter obtained based on STEP21 finely tunes the parameter of whole multilayered model further, and this step is one prison Supervise and instruct experienced process；STEP21 is similar to the random initializtion initial value process of neutral net, due to the degree of depth study STEP21 be not with Machine initializes, but obtained by the structure of study input data, thus this initial value is closer to global optimum such that it is able to Obtain more preferable effect；So degree of deep learning effect quality largely gives the credit to the feature learning process of STEP21；

For the vehicle image data set of tape label, the present invention uses web crawlers technology to collect the vehicle figure of various vehicles Picture, the vehicle image so collected again through manual confirmation as the vehicle image data of tape label；

About wake-sleep algorithm, seeing accompanying drawing 8, its main thought is to given generation weights in the wake stage, passes through Study obtains cognitive weights；In the sleep stage to given cognitive weights, obtain generating weights by study；

In the wake stage, l layer generates weights g_l, it is updated with formula (9),

Δg_l=ε s^l+1(s^l-p^l) (9)

In formula, Δ g_lIt is that l layer generates weights g_lRenewal changing value, ε is learning rate, s^l+1It is l+1 layer neuron Liveness, s^lIt is l layer neuron liveness, p^lIt it is the activation probability during current state driving of l layer neuron；

The sleep stage, l layer cognition weight w_l, it is updated with formula (10),

Δw_l=ε s^l-1(s^l-q^l) (10)

In formula, Δ w_lIt is l layer cognition weight w_lRenewal changing value, ε is learning rate, s^l-1It is l-1 layer neuron Liveness, s^lIt is l layer neuron liveness, q^lIt is that l layer neuron is used the preceding layer neuron of present cognitive weights Activation probability when current state drives；

(3) design the grader of a kind of type subclass being applicable to thousands of kinds, and there is autgmentability；

The present invention uses Softmax grader to classify vehicle；Softmax grader is in the situation of feature robust Under, there is preferable classifying quality, the most this grader has autgmentability, without to original training after new vehicle occurs Good network characterization carries out relearning training, adds the practicality of system；Softmax principle is the input number that will extract Comparing with feature database according to feature, calculate the probability of each classification results, the result then taking probability the highest is entered Row output；

Using the learning outcome in CNN as the input data of softmax grader；It is to divide towards multiclass that Softmax returns The Logistic of class problem returns, and is the general type of Logistic recurrence, it is adaptable to the situation of mutual exclusion between classification.It is right to assume In training set { (x⁽¹⁾,y⁽¹⁾,…,x^(m),y^(m)), there is y⁽¹⁾∈ 1,2 ..., and k}, for given sample input x, export one The vector of k dimension represents that the probability that each classification results occurs is p (y=i | x), it is assumed that function h (x) is as follows:

$h_{\mathrm{\θ}}\left(x^{\left(i\right)}\right)=\left[\begin{array}{c}p(y^{\left(i\right)}=1|x^{\left(i\right)},\mathrm{\θ})\\ p(y^{\left(i\right)}=1|x^{\left(i\right)},\mathrm{\θ})\\ .\\ .\\ .\\ p(y^{\left(i\right)}=k|x^{\left(i\right)},\mathrm{\θ})\end{array}\right]=\frac{1}{{\Σ}_{j=1}^k{e}^{{\mathrm{\θ}}_j^T{x}^{\left(i\right)}}}\left[\begin{array}{c}{e}^{{\mathrm{\θ}}_1^T{x}^{\left(i\right)}}\\ e^{{\mathrm{\θ}}_2^T{x}^{\left(i\right)}}\\ .\\ .\\ .\\ e^{{\mathrm{\θ}}_k^T{x}^{\left(i\right)}}\end{array}\right]---\left(11\right)$

θ₁,θ₂,…θ_kIt is the parameter of model, and all of probability and be 1.Adding the cost function after regularization term is:

$J\left(\mathrm{\θ}\right)=-\frac{1}{m}\[\underset{i=1}{\overset{m}{\Σ}}\underset{j=1}{\overset{k}{\Σ}}1\{y^{\left(i\right)}=j\}\log \frac{e^{{\mathrm{\θ}}_j^T{x}^{\left(i\right)}}}{{\Σ}_{l=1}^k{e}^{{\mathrm{\θ}}_l^T{x}^{\left(i\right)}}}\]+\frac{\mathrm{\λ}}{2}\underset{l=1}{\overset{k}{\Σ}}\underset{j=0}{\overset{n}{\Σ}}{\mathrm{\θ}}_{ij}^2---\left(12\right)$

The partial derivative of l parameter of jth classification is by cost function:

${\▿}_{{\mathrm{\θ}}_j}J\left(\mathrm{\θ}\right)=-\frac{1}{m}\underset{i=1}{\overset{m}{\Σ}}\[x^{\left(i\right)}\left(1\right\{y^{\left(i\right)}=j\}-p(y^{\left(i\right)}=j|x^{\left(i\right)};\mathrm{\θ}\left)\right)\}\]+{\mathrm{\λ\θ}}_j---\left(13\right)$

Finally, by minimizing J (θ), it is achieved the classification of softmax returns.

(4) design one quick R-CNN real-time performance vehicle detection end to end of one use truly and knowledge Other framework；

The present invention to solve emphatically three below problem:

1) how design section advises network；

2) region how is trained to advise network；

3) region suggestion network and quick R-CNN network sharing features how is allowed to extract network；

Region suggestion Web content specifically includes that region suggestion network structure, feature extraction, the design of region suggestion network With training thinking, candidate region, window classification and position refine；

Region suggestion network structure, feature extraction, the design of region suggestion network:

The design of network is advised as shown in Figure 11, in order to enable multiple in region suggestion network structure, feature extraction, region Calculating on GPU, be divided into some group in calculating in layer, often the calculating in group is completed by its corresponding GPU, so can promote Calculate speed；A kind of VGG network shown in Figure 12, i.e. visual geometry group, 8 Ge Juan basic units, 3 full articulamentums, Amount to 11 layers；8 Ge Juan basic units there are 5 convolutional layers organized, 2 classification layers extract characteristics of image, 1 classification layer characteristic of division；3 Individual full articulamentum link sort layer 6 respectively, classification layer 7 and classification layer 8；

In Figure 11, normalized 224 × 224 images are sent directly into network, and first five stage is the convolution+ReLU+ pond on basis The form changed, inputs p candidate region at the ending of the 5th stage again, and candidate region is with 1 picture numbers and 4 geometry positions Confidence ceases；Each candidate region is uniformly divided into M × N block by the RoI-pond layer at the ending of the 5th stage, carries out every piece Great Chiization operates；Candidate region not of uniform size on characteristic pattern is changed into the data that size is unified, send into next layer training and Identify；First five stage technique above-mentioned is all the technology being proved to maturation in convolutional neural networks technology, it is important to how to make this A little candidate regions can be with the network characterization in these first five stages of picture of multiplexing；

Region suggestion network and quick R-CNN network sharing features extraction network:

Accompanying drawing 6 describes the convolutional layer output that region advises how network and quick R-CNN share, first five stage in Figure 11 Belong to a primitive character and extract network, to region suggestion network formation zone Suggestion box and examine to quick R-CNN the most respectively Survey the characteristics of image in the Suggestion box of region, be then output to two full layers connected at the same level, i.e. classification layer 6+ in Figure 11 ReLU6 and classification layer 7+ReLU7, obtains p the characteristic pattern containing 4096 vectors, gives classification layer the most respectively and window returns Layer processes；

Formation zone Suggestion box:

For formation zone Suggestion box, the present invention is sliding in the convolution Feature Mapping of last convolutional layer shared output Dynamic little network, this network is connected to input in the spatial window of the n × n of convolution Feature Mapping, as shown in Figure 5 entirely；Each Sliding window is mapped on a low dimensional vector, and the low dimensional vector in accompanying drawing 5 is 256-d, a slip of each Feature Mapping The corresponding numerical value of window；This vector output is to two full layers connected at the same level；Individual window returns layer and a classification Layer；Window returns layer and exports on each position, recommends region correspondence window to need have translation scaling invariance, returns for 9 kinds Layer exports 4 translation zooming parameters from 256 dimensional features；Classification layer exports from 256 dimensional features and belongs to the general of foreground and background Rate；

Such as, in the position of each sliding window, k region suggestion of prediction simultaneously, therefore window returns layer 4k Output, i.e. the codes co-ordinates of k Suggestion box；Classification layer output 2k Suggestion box score, to each Suggestion box be i.e. vehicle target/ The estimated probability of non-vehicle target；

Candidate region: the frame parametrization that k Suggestion box is referred to as candidate region by corresponding k, the most each candidate region with Centered by current sliding window mouth center, and corresponding a kind of yardstick and length-width ratio, the present invention uses 3 kinds of yardsticks and 3 kinds of length-width ratios, as Shown in accompanying drawing 3；So just there is k=9 kind candidate region at each sliding position；The convolution feature that size is W × H is reflected Penetrate, a total of W × H × k candidate region；

In whole quick R-CNN algorithm, have three kinds of graphical rules: 1) artwork yardstick: the size of original input picture, Unrestricted, do not affect performance；2) Normalized Scale: input feature vector extracts the size of network, is arranged when test, candidate Region sets on this yardstick；The relative size of this parameter and candidate region determines the target zone wanting detection；3) Network input yardstick: the size of input feature vector detection network, arranges when training, is 224 × 224.

In sum, region suggestion network is as input using an image, exports the set of rectangular target Suggestion box, often Individual frame has the score of a Vehicle Object, as shown in Figure 7；

Suggestion Web content in training region specifically includes that training sample, cost function and hyper parameter；

The training of region suggestion network:

The training of region suggestion network, the most whether the present invention distributes a binary label to each candidate region, is Vehicle Object；Here positive label is distributed to two class candidate regions: (i) and certain GT, the enclosing region of ground truth has The ratio of high common factor union, IoU, Intersection-over-Union, overlapping candidate region；(ii) surround with any GT The candidate region that the region IoU more than 0.7 is overlapping；The negative label of distribution simultaneously is given the lowest with the IoU ratio of all GT enclosing region In the candidate region of 0.3；Leave out the candidate region of anon-normal non-negative；

Training sample algorithm:

STEP31: order reads every figure in training set；

STEP32: the true value candidate region to each demarcation, the candidate region overlapping ratio maximum is designated as prospect sample This；

STEP33:: candidate region remaining to STEP32, if it is overlapping with certain demarcation, IoU ratio is more than 0.7, note For prospect sample；If its overlap proportion demarcated with any one is both less than 0.3, it is designated as background sample；

STEP34: candidate region remaining to STEP32 and STEP33 is discarded；

STEP35: the candidate region crossing over image boundary is discarded.

Cost function:

Define according to these, the multitask cost followed here, use and minimize object function；Cost to an image Function formula (14) represents,

$L\left(\right\{p_i\},\{t_i\left\}\right)=\frac{1}{N_{cls}}\underset{i}{\Σ}{L}_{cls}(p_i,p_i^{*})+\mathrm{\λ}\frac{1}{N_{reg}}\underset{i}{\Σ}{p}_i^{*}{L}_{reg}(t_i,t_i^{*})---\left(14\right)$

In formula, i is the index of candidate region, N in a batch processing_clsFor the normalization coefficient of layer of classifying, N_regFor returning The normalization coefficient of layer, λ is balance weight, p_iFor the prediction probability of vehicle target,For GT label, if candidate region is justIf candidate region is negativet_iIt is a vector, represents 4 the parametrization coordinates surrounding frame of prediction,For The GT corresponding with positive candidate region surrounds the coordinate vector of frame, L_clsFor the logarithm cost of classification, L_regFor returning logarithm cost, L ({p_i},{t_i) it is total logarithm cost；

Logarithm cost L of classification_clsCalculated by formula (15),

$L_{cls}(p_i,p_i^{*})=-log\[p_i^{*}{p}_i+(1-p_i^{*})(1-p_i)\]---\left(15\right)$

Window returns logarithm cost L_regCalculated by formula (16),

$L_{reg}(t_i,t_i^{*})=R(t_i-t_i^{*})---\left(16\right)$

In formula, R is the cost function of the robust of definition, belongs to Smooth L1 error, insensitive to outlier, uses formula (17) calculate,

In formula (14)This means only positive candidate region, i.e.Shi Caiyou returns generation Valency, other situations due toDo not return cost；Classification layer and window return the output of layer respectively by { p_iAnd { t_iGroup Becoming, these two respectively by N_clsAnd N_regAnd a balance weight λ normalization, select λ=10, N here_cls=256, N_reg= 2400, returning layer item by such selection sort layer and window is almost equal weight；

About position refine, using 4 values, centre coordinate, width and height here, computational methods are as follows,

$\begin{array}{c}{t}_x=(x-x_a)/w_a,t_y=(y-y_a)/h_a,t_w=\log (w/w_a),t_h=\log (h/h_a),\\ {t^{*}}_x=(x^{*}-x_a)/w_a,{t^{*}}_y=(y^{*}-y_a)/h_a,{t^{*}}_w=\log (w^{*}/w_a),{t^{*}}_h=\log (h^{*}/h_a)\end{array}---\left(18\right)$

In formula, x, y, w, h represent encirclement frame centre coordinate, width and height, x respectively_a、y_a、w_a、h_aRepresent candidate respectively Regional center coordinate, width and height, x^*、y^*、w^*、h^*Represent the encirclement frame centre coordinate of prediction, width and height respectively；With The result of calculation of formula (18) carries out position refine；It practice, explicitly do not extract any candidate window, use district completely Territory suggestion network self completes to judge and position refine；

The optimization of region suggestion network:

Region suggestion network can be embodied as full convolutional network naturally, end-to-end by back propagation and stochastic gradient descent Training；Here the sampling policy using picture centre trains this network, and each batch processing is by containing many positive negative samples Single image forms；256 candidate regions of sampling the most in one image, calculate the cost in batch processing with formula (14) Function, the ratio of the positive and negative candidate region wherein sampled is 1:1；If the positive sample number in an image is less than 128, we are just Other residue candidate regions in this batch processing are filled up with negative sample；Here batch processing is dimensioned to 256；

The all new layers of random initializtion, institute is come by the weight obtained from the Gauss distribution that zero-mean standard deviation is 0.01 Call new layer and refer to the layer after region suggestion network, such as the classification layer 6+ReLU6 in accompanying drawing 11 and classification layer 7+ReLU7； Every other layer, the convolutional layer i.e. shared, such as first five layer in accompanying drawing 11, by the classification samples pre-training to vehicle vehicle Model comes initialized；The present invention of learning rate to(for) 60k batch processing on vehicle model data collection is 0.001, for The learning rate of next 20k batch processing is 0.0001；Momentum is 0.9, and weight decays to 0.0005；

The model initialization of vehicle vehicle pre-training: be broadly divided into data prepare, calculate image average, the definition of network, 5 steps such as training and recovery data；

1) data prepare；New folder myself in data, we have collected all kinds of vehicle by reptile software View data, due to scan for obtaining with keyword substantially with the vehicle image data of label, we are by it As training data；Another kind of data are the vehicle image data that we are obtained by bayonet camera；

Input train.txt and val.txt of training and test describes, and lists All Files and their mark Sign；The name of classification is the order of ASCII character, i.e. 0-999, and corresponding systematic name is mapped in synset_ with numeral In words.txt；Val.txt can not label, be all arranged to 0；Then the size of picture is unified into 256 × 256；Then exist Newly-built myself file in caffe-master/examples, then by caffe-maester/examples/imagenet Create_imagenet.sh copy to, under this document folder, its name change create_animal.sh into, amendment training and surveying The setting in examination path, runs this sh；Finally obtain myself_train_lmdb and myself_val_lmdb；

2) image average is calculated；Model needs us to deduct average from every pictures, so we must obtain training Average, realizes with tools/compute_image_mean.cpp, replicates caffe-maester/examples/ equally In ./make_imagenet_mean to the examples/myself of imagenet, it is renamed as make_car_mean.sh, Revised path；

3) definition of network；All Files in caffe-master/models/bvlc_reference_caffenet Copy in caffe-master/examples/myself file, revise train_val.prototxt, note revising number Path according to layer；

In training, we are with a softmax loss layer counting loss function and initialize back propagation, and are testing Card, our service precision layer detects our precision；Also having agreement solver.prototxt run, duplication comes, will The first row path changes our path net: " examples/myself/train_val.prototxt " into,

Test_iter:1000 refers to the batch of test；Test_interval:1000 refers to every 1000 iteration tests one Secondary；Base_lr:0.01 is basic learning rate；Lr_policy: " step " learning rate changes；The change of gamma:0.1 learning rate Ratio；Every 100000 iteration of stepsize:100000 reduce learning rate；The every 20 layers of display of display:20 are once；max_ Iter:450000 maximum iteration time；The parameter of momentum:0.9 study；The parameter of weight_decay:0.0005 study； 10000 display states of the every iteration of snapshot:10000；Solver_mode:GPU end adds a line, represents and uses GPU computing；

4) training；Train_caffenet.sh in caffe-master/examples/imagenet is replicated And revise entitled train_myself.sh operation, the path inside amendment；

5) data are recovered；Resume_training.sh in caffe-master/examples/imagenet is replicated Come over and run；

The model initialization work of vehicle vehicle pre-training is completed through above-mentioned process；Further, the present invention proposes one Plant 4 step training algorithms, learnt the feature shared by alternative optimization；

STEP41: with region suggestion network optimization training region suggestion network, with above-mentioned data prepare, calculating image equal Value, the definition of network, train and recover 5 steps such as data and complete model initialization, and end-to-end fine setting is advised for region Task；

STEP42: the Suggestion box generated with the region suggestion network of STEP1, is trained a single inspection by quick R-CNN Survey grid network, this detection network is by the model initialization of vehicle vehicle pre-training equally, and at this time two networks also do not have Shared volume lamination；

STEP43: with detection netinit region suggestion network training, fix the convolutional layer shared, and only finely tune district The layer that territory suggestion network is exclusive, at this moment two network shared volume laminations；

STEP44: keep the convolutional layer shared to fix, finely tune the classification layer of quick R-CNN；So, two networks share phase Same convolutional layer, finally constitutes a unified network；

The visual identity of vehicle vehicle:

We provide a vehicle vehicle visual identity main flow below, and whole handling process is as shown in Figure 13；

STEP51: read image to be identified；

STEP52: be normalized image to be identified, obtains tri-different colours 224 × 224 of RGB normalized View data；

STEP53: tri-normalized view data of different colours 224 × 224 of RGB are input to three CNN passages, through 5 Layer process of convolution obtains vehicle vehicle character image data；

STEP54: the Suggestion box generated vehicle vehicle character image data by region suggestion network, chooses one The Suggestion box of high score, i.e. obtains an area-of-interest, RoI；By maximum 5 layers of pondization, this RoI is carried out process obtain The trellis diagram of one 6 × 6 × 256RoI；

STEP55: the trellis diagram of RoI is exported the feature obtaining 4096 dimensions to two full layers connected at the same level after processing Vector, as the input data of softmax grader；

STEP56: the classification regression analysis to characteristic vector softmax of 4096 dimensions obtains vehicle vehicle cab recognition result, Vehicle vehicle classification is become 1000 kinds by the present invention, thus identifies by which kind of vehicle the vehicle in altimetric image belongs to.

Embodiment 2

The Visual identification technology of the present invention has universality, it is adaptable to the subclass identification of other objects, as long as participating in training Data run learn in the system that the present invention develops, it is thus achieved that can be achieved with this class object after the feature of this class object Subclass identification mission.

Embodiment 3

The Visual identification technology of the present invention has autgmentability, without to original network trained after new subclass occurs Feature carries out relearning training, as long as new subclass is trained study, and softmax grader in systems expands The data of exhibition classification.

The foregoing is only the preferable implementation example of the present invention, be not limited to the present invention, all in present invention spirit and Within principle, any modification, equivalent substitution and improvement etc. made, should be included within the scope of the present invention.

Claims (10)

1. a model recognizing method based on quick R-CNN deep neural network, it is characterised in that: include one for the degree of depth The VGG network that study and training identify, a region suggestion network being used for extracting area-of-interest and one are for vehicle The Softmax grader of classification；

Described VGG network, including 8 Ge Juan basic units, 3 full articulamentums, 11 layers altogether；8 Ge Juan basic units have 5 convolution organized Layer, 2 classification layers extract characteristics of image, 1 classification layer characteristic of division；3 full articulamentum link sort layers 6 respectively, classification layers 7 With classification layer 8；

Described region suggestion network, returns layer, 1 module calculating Classification Loss and 1 including 1 classification layer, 1 window Calculation window returns the module of loss, p Suggestion box interested of output；

Described Softmax grader, obtains feature database data by the input data characteristics and the learning training that extract and compares Right, calculate the probability of each classification results, the result then taking probability the highest exports；

Quickly R-CNN deep neural network, has accessed described region suggestion network at the 5th layer of end of described VGG network, Described region is advised, and network shares low-level image feature extraction process and the result of first 5 layers of described VGG network；

The 6th layer of described VGG network and the 7th layer p the Suggestion box interested according to described region suggestion network output Interior characteristics of image carries out convolution and ReLU process, obtains p the characteristic pattern containing 4096 vectors, gives classification layer the most respectively Return layer with window to process, it is achieved the segmentation of vehicle image；On the other hand, p is contained by described Softmax grader The characteristic pattern of 4096 vectors carries out Classification and Identification, obtains the classification results of vehicle vehicle.

2. model recognizing method as claimed in claim 1, it is characterised in that: described Softmax grader, at learning training Period using the learning outcome in quick R-CNN deep neural network as the input data of softmax grader；Softmax returns Returning is that the Logistic towards multicategory classification problem returns, and is the general type of Logistic recurrence, it is adaptable between classification mutually Situation about scolding；Assume for training set { (x⁽¹⁾,y⁽¹⁾,…,x^(m),y^(m)), there is y⁽¹⁾∈ 1,2 ..., k}, for given sample This input x, exports the vector of k dimension and represents that the probability that each classification results occurs is p (y=i | x), it is assumed that function h X () is as follows:

$h_{\mathrm{\θ}}\left(x^{\left(i\right)}\right)=\left[\begin{array}{c}p(y^{\left(i\right)}=1|x^{\left(i\right)},\mathrm{\θ})\\ p(y^{\left(i\right)}=1|x^{\left(i\right)},\mathrm{\θ})\\ .\\ .\\ .\\ p(y^{\left(i\right)}=k|x^{\left(i\right)},\mathrm{\θ})\end{array}\right]=\frac{1}{{\Σ}_{j=1}^k{e}^{{\mathrm{\θ}}_j^T{x}^{\left(i\right)}}}\left[\begin{array}{c}{e}^{{\mathrm{\θ}}_1^T{x}^{\left(i\right)}}\\ e^{{\mathrm{\θ}}_2^T{x}^{\left(i\right)}}\\ .\\ .\\ .\\ e^{{\mathrm{\θ}}_k^T{x}^{\left(i\right)}}\end{array}\right]---\left(11\right)$

θ₁,θ₂,…θ_kIt is the parameter of model, and all of probability and be 1；Adding the cost function after regularization term is:

$J\left(\mathrm{\θ}\right)=-\frac{1}{m}\[\underset{i=1}{\overset{m}{\Σ}}\underset{j=1}{\overset{k}{\Σ}}1\{y^{\left(i\right)}=j\}\log \frac{e^{{\mathrm{\θ}}_j^T{x}^{\left(i\right)}}}{{\Σ}_{l=1}^k{e}^{{\mathrm{\θ}}_l^T{x}^{\left(i\right)}}}\]+\frac{\mathrm{\λ}}{2}\underset{l=1}{\overset{k}{\Σ}}\underset{j=0}{\overset{n}{\Σ}}{\mathrm{\θ}}_{ij}^2---\left(12\right)$

The partial derivative of l parameter of jth classification is by cost function:

${\▿}_{{\mathrm{\θ}}_j}J\left(\mathrm{\θ}\right)=-\frac{1}{m}\underset{i=1}{\overset{m}{\Σ}}\[x^{\left(i\right)}\left(1\right\{y^{\left(i\right)}=j\}-p(y^{\left(i\right)}=j|x^{\left(i\right)};\mathrm{\θ}\left)\right)\]+{\mathrm{\λ\θ}}_j---\left(13\right)$

In formula, j is classification number, and m is the classification number of training set, p (y⁽ⁱ⁾=j | x⁽ⁱ⁾；Being θ)) } the x probability that is divided into classification j, λ is Regularization term parameter, also referred to as weight attenuation term, this regularization term parameter mainly prevents over-fitting；

Finally, by minimizing J (θ), it is achieved the classification of softmax returns, and is saved in feature database by classification regression result；

When identifying classification, the input data characteristics and the learning training that extract are obtained feature database data and compares, calculate Going out the probability of each classification results, the result then taking probability the highest exports.

3. model recognizing method as claimed in claim 1, it is characterised in that: described region suggestion network, in order to generate district Territory Suggestion box, has accessed described region suggestion network, i.e. at the convolutional layer of the 5th layer at the 5th layer of end of described VGG network Slide on the convolution Feature Mapping figure of output a little network, and this network is connected to the n × n's of input convolution Feature Mapping entirely In spatial window；Each sliding window is mapped on a low dimensional vector, and low dimensional vector is 256-d, the one of each Feature Mapping The corresponding numerical value of individual sliding window；This vector output is to two full layers connected at the same level；Individual window returns layer and Individual classification layer；Window returns layer and exports on each position, recommends region correspondence window to need have translation for 9 kinds and scales constant Property, window returns layer and exports 4 translation zooming parameters from 256 dimensional features, has 4k output, i.e. the coordinate of k Suggestion box is compiled Code；Classification layer exports the probability belonging to foreground and background from 256 dimensional features, exports 2k Suggestion box score, i.e. builds each View frame is the estimated probability of vehicle target/non-vehicle target.

4. the model recognizing method as described in claim 1 or 3, it is characterised in that: the training of region suggestion network, to each time The most whether favored area one binary label of distribution, be Vehicle Object；Here positive label is distributed to two class candidate regions:

I () and the enclosing region of certain GT have the ratio of the highest common factor union, IoU, overlapping candidate region；

(ii) candidate region overlapped with any GT enclosing region IoU more than 0.7；The negative label of distribution simultaneously is given and all GT bags The IoU ratio enclosing region is below the candidate region of 0.3；

Leave out the candidate region of anon-normal non-negative；Specific algorithm is as follows:

STEP31: order reads every figure in training set；

STEP32: the true value candidate region to each demarcation, the candidate region overlapping ratio maximum is designated as prospect sample；

STEP33:: candidate region remaining to STEP32, if it is overlapping with certain demarcation, IoU ratio is more than 0.7, before being designated as Scape sample；If its overlap proportion demarcated with any one is both less than 0.3, it is designated as background sample；

STEP34: candidate region remaining to STEP32 and STEP33 is discarded；

STEP35: the candidate region crossing over image boundary is discarded.

5. the model recognizing method as described in claim or 4, it is characterised in that: in order to automatically carry out candidate region screening and Regional location refine, uses here and minimizes object function；The cost function of one image is represented with formula (14),

$L\left(\right\{p_i\},\{t_i\left\}\right)=\frac{1}{N_{cls}}\underset{i}{\Σ}{L}_{cls}(p_i,p_i^{*})+\mathrm{\λ}\frac{1}{N_{reg}}\underset{i}{\Σ}{p}_i^{*}{L}_{reg}(t_i,t_i^{*})---\left(14\right)$

In formula, i is the index of candidate region, N in a batch processing_clsFor the normalization coefficient of layer of classifying, N_regReturn for window The normalization coefficient of layer, λ is balance weight, p_iFor the prediction probability of vehicle target,For GT label, if candidate region is justIf candidate region is negativet_iIt is a vector, represents 4 the parametrization coordinates surrounding frame of prediction,For The GT corresponding with positive candidate region surrounds the coordinate vector of frame, L_clsFor the logarithm cost of classification, L_regFor returning logarithm cost, L ({p_i},{t_i) it is total logarithm cost；

Logarithm cost L of classification_clsCalculated by formula (15),

$L_{cls}(p_i,p_i^{*})=-log\[p_i^{*}{p}_i+(1-p_i^{*})(1-p_i)\]---\left(15\right)$

Window returns logarithm cost L_regCalculated by formula (16),

$L_{reg}(t_i,t_i^{*})=R(t_i-t_i^{*})---\left(16\right)$

In formula, R is the cost function of the robust of definition, belongs to Smooth L1 error, insensitive to outlier, with formula (17) Calculate,

In formula (14)This means only positive candidate region, i.e.Shi Caiyou returns cost, its His situation due toDo not return cost；Classification layer and window return the output of layer respectively by { p_iAnd { t_iComposition, these are two years old Item is respectively by N_clsAnd N_regAnd a balance weight λ normalization, select λ=10, N here_cls=256, N_reg=2400, pass through It is almost equal weight that such selection sort layer and window return layer item；

About position refine, using 4 values, centre coordinate, width and height here, computational methods are as follows,

$\begin{array}{c}{t}_x=(x-x_a)/w_a,t_y=(y-y_a)/h_a,t_w=\log (w/w_a),t_h=\log (h/h_a),\\ {t^{*}}_x=(x^{*}-x_a)/w_a,{t^{*}}_y=(y^{*}-y_a)/h_a,{t^{*}}_w=\log (w^{*}/w_a),{t^{*}}_h=\log (h^{*}/h_a)\end{array}---\left(18\right)$

In formula, x, y, w, h represent encirclement frame centre coordinate, width and height, x respectively_a、y_a、w_a、h_aRepresent respectively in candidate region Heart coordinate, width and height, x^*、y^*、w^*、h^*Represent the encirclement frame centre coordinate of prediction, width and height respectively；Use formula (18) result of calculation carries out refine region, position suggestion network.

6. model recognizing method as claimed in claim 1, it is characterised in that: described VGG network, at label vehicle image number The method setting up multilayer neural network on according to, is divided into two steps, and one is every time training one layer network, and two is tuning, makes original representation X The X' that the senior expression r upwards generated generates downwards with this senior expression r is the most consistent；

The propagated forward process of convolutional neural networks, the output of last layer is the input of current layer, and by activation primitive by Layer transmission, Practical Calculation output formula (4) of the most whole network represents,

O_p=F_n(…(F₂(F₁(XW₁)W₂)…)W_n) (4)

In formula, X represents and is originally inputted, F_lRepresent the activation primitive of l layer, W_lRepresent the mapping weight matrix of l layer, O_pRepresent The Practical Calculation output of whole network；

The output of current layer formula (5) represents,

X^l=f^l(W^lX^l-1+b^l) (5)

In formula, l represents the network number of plies, X^lRepresent the output of current layer, X^l-1The output of expression last layer, i.e. the input of current layer, W^lRepresent trained, the mapping weight matrix of current network layer, b^lAdditivity for current network is bigoted, f^lIt it is current net The activation primitive of network layers；The activation primitive f used^lFor correcting linear unit, i.e. ReLU, represent with formula (6),

$f^l=max({\left(W^l\right)}^T{X}^l,0)=\left\{\begin{array}{cc}{\left(W^l\right)}^T{X}^l& {\left(W^l\right)}^T{X}^l>0\\ 0& {\left(W^l\right)}^T{X}^l\≤0\end{array}\right.---\left(6\right)$

In formula, l represents the network number of plies, W^lRepresent trained, the mapping weight matrix of current network layer, f^lIt it is current net The activation primitive of network layers；Its effect is that then allowing it is 0 if convolutional calculation result is less than 0；Otherwise keep its value constant.

7. the model recognizing method as described in claim 1 or 6, it is characterised in that: described VGG network, first 5 layers is an allusion quotation The degree of depth convolutional neural networks of type, this neural metwork training is a back-propagation process, by error function back propagation, profit By stochastic gradient descent method, deconvolution parameter and biasing are optimized and revised, until network convergence or reach maximum iteration time Stop；

Back propagation needs by comparing the training sample with label, uses square error cost function, for c Classification, the multi-class of N number of training sample is identified, and network final output error function formula (7) calculates error,

$E^N=\frac{1}{2}\underset{n=1}{\overset{N}{\Σ}}\underset{k=1}{\overset{c}{\Σ}}{(t_k^n-y_k^n)}^2---\left(7\right)$

In formula, E^NFor square error cost function,It is the kth dimension of the n-th sample corresponding label,It it is the n-th sample correspondence net The kth output of network prediction；

When error function is carried out back propagation, use computational methods as the BP class of algorithms, as shown in formula (8),

$\begin{array}{c}{\mathrm{\δ}}^l={\left(W^{l+1}\right)}^T{\mathrm{\δ}}^{l+1}\×f^{\′}\left(u^l\right)\\ u^l=W^l{x}^{l-1}+b^l\end{array}---\left(8\right)$

In formula, δ^lRepresent the error function of current layer, δ^l+1Represent the error function of last layer, W^l+1For last layer mapping matrix, f' Represent the inverse function of activation primitive, i.e. up-sample, u^lRepresent the output not by the last layer of activation primitive, x^l-1Represent next The input of layer, W^lWeight matrix is mapped for this layer；

After error back propagation, obtain the error function δ of each Internet^l, then use stochastic gradient descent method to network Weights W^lModify, then carry out next iteration, until network reaches the condition of convergence；Need when carrying out error propagation first to lead to Crossing the up-sampling in formula (8) makes before and after's two-layer equivalently-sized, then carries out error propagation；

Algorithm idea is: 1) the most successively build monolayer neuronal unit, is one single layer network of training the most every time；2) when all After layer has been trained, wake-sleep algorithm is used to carry out tuning；

Degree of deep learning training process is specific as follows:

STEP21: use unsupervised learning from bottom to top, i.e. from the beginning of bottom, past top layer training in layer, learn car Characteristics of image: first with without label vehicle image data training ground floor, first learn the parameter of ground floor during training, due to model Hold quantitative limitation and sparsity constraints so that the model obtained can learn the structure to data itself, thus it is defeated to obtain ratio Enter to have more the feature of expression ability；After study obtains l-1 layer, using the output of l-1 layer as the input of l layer, train L layer, thus respectively obtains the parameter of each layer；

STEP22: top-down supervised learning, i.e. goes training, the top-down biography of error by the vehicle image data of tape label Defeated, network is finely adjusted.

8. model recognizing method as claimed in claim 1, it is characterised in that: the model of first 5 layers of described VGG network is initial Change: be broadly divided into data and prepare, calculate image average, the definition of network, train and recover 5 steps of data；

1) data prepare；Collecting the view data of all kinds of vehicle, obtain is substantially the vehicle image data with label, will It is as training image data；Another kind of data are the vehicle image data obtained by bayonet camera；

2) image average is calculated；Average is deducted from every pictures；

3) definition of network；Main definitions xml tag path, the path of picture, deposit train.txt, val.txt, The path of test.txt and trainval.txt file；

4) training；Run training module；

5) data are recovered；The layer of ReLu5 before deleting, and change the bottom of roi_pool5 into data and rois；

The model initialization work of vehicle vehicle pre-training is completed through above-mentioned process.

9. the model recognizing method as described in claim 1 or 8, it is characterised in that: network utilisation institute is advised in described region Front 5 layers of low-level image feature of the VGG network stated extract result, and i.e. two networks have shared front 5 layers of bottom spy of described VGG network Levy, need to learn to optimize the feature shared by alternative optimization；Specific algorithm is as follows:

STEP41: with the optimization training region suggestion network of region suggestion network, with described data prepare, calculating image equal Value, the definition of network, train and recover 5 steps of data and complete model initialization, and end-to-end fine setting for region suggestion appoint Business；

STEP42: the Suggestion box generated with the region suggestion network of STEP1, is trained a single detection net by quick R-CNN Network, this detection network is by the model initialization of vehicle vehicle pre-training equally, and at this time two networks are not the most shared Convolutional layer；

STEP43: with detection netinit region suggestion network training, fix the convolutional layer shared, and only fine setting region is built The layer that view network is exclusive, at this moment two network shared volume laminations；

STEP44: keep the convolutional layer shared to fix, finely tune the classification layer of quick R-CNN；So, two networks are shared identical Convolutional layer, finally constitutes a unified network.

10. model recognizing method as claimed in claim 1, it is characterised in that: vehicle vehicle visual identity main flow is as follows；

STEP51: read image to be identified；

STEP52: be normalized image to be identified, obtains tri-normalized images of different colours 224 × 224 of RGB Data；

STEP53: tri-normalized view data of different colours 224 × 224 of RGB are input to three CNN passages, through 5 layers of volume Long-pending process obtains vehicle vehicle character image data；

STEP54: the Suggestion box that vehicle vehicle character image data is generated by region suggestion network, choose one the highest The Suggestion box divided, i.e. obtains an area-of-interest, RoI；This RoI is carried out process by maximum 5 layers of pondization and obtains one 6 The trellis diagram of × 6 × 256RoI；

STEP55: the trellis diagram of RoI is exported the characteristic vector obtaining 4096 dimensions to two full layers connected at the same level after processing, Input data as softmax grader；

STEP56: the classification regression analysis to characteristic vector softmax of 4096 dimensions obtains vehicle vehicle cab recognition result, identifies Go out by which kind of vehicle the vehicle in altimetric image belongs to.