CN106469304A - Handwritten signature location positioning method in bill based on depth convolutional neural networks - Google Patents
Handwritten signature location positioning method in bill based on depth convolutional neural networks Download PDFInfo
- Publication number
- CN106469304A CN106469304A CN201610841643.8A CN201610841643A CN106469304A CN 106469304 A CN106469304 A CN 106469304A CN 201610841643 A CN201610841643 A CN 201610841643A CN 106469304 A CN106469304 A CN 106469304A
- Authority
- CN
- China
- Prior art keywords
- bill
- convolutional neural
- neural networks
- handwritten signature
- layer
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/30—Writer recognition; Reading and verifying signatures
- G06V40/33—Writer recognition; Reading and verifying signatures based only on signature image, e.g. static signature recognition
Landscapes
- Engineering & Computer Science (AREA)
- Human Computer Interaction (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses handwritten signature location positioning method in bill based on depth convolutional neural networks, specifically implement according to following steps：Step 1：Build the platform based on the caffe deep learning framework comprising multiple convolutional neural networks models；Step 2：Prepare the data set of bill；Step 3：Training network obtains detection and localization model；Step 4：The detection and localization model being obtained with the step 3 signature position hand-written to bill positioning to be detected.Handwritten signature location positioning method in the bill based on depth convolutional neural networks for the present invention, can accurately calibrate the position containing handwritten signature in bill.
Description
Technical field
The invention belongs to framing detection technique field is and in particular to a kind of bill based on depth convolutional neural networks
Middle handwritten signature location positioning method.
Background technology
At present, China to the research of bill automatic checkout system mostly also in the development phase, practical application is less, and joins
Putting a set of bill specifies handwritten signature position detecting system relatively costly, also limit the development of bill test technique automatic.
This also just proposes demand to vast research institution and scholar, needs to work out a kind of technology of bill Aulomatizeted Detect.Ticket
According in handwritten signature there is the characteristic of character, the identification for character mainly has：Statistical nature technology of identification, architectural feature are known
Other technology, the technology of identification based on neutral net.Statistical nature has character two dimensional surface position feature, character horizontally or vertically
Histogram feature of direction projection etc., the character recognition technologies based on statistical nature are weak for nearly word form symbol separating capacity, only suitable
Close character rough sort.Architectural feature has the trend of stroke, isolated point and whether contains closure stroke etc., and this method is easy to
Distinguish font change character greatly.At present, is in the research of neutral net a new peak time, neutral net is
It is widely used in area of pattern recognition.With academia for deep learning research, the algorithm of deep learning is more and more ripe,
Application also gets more and more.But, most of neutral nets are only used for extracting clarification of objective it is impossible to position for target.
Content of the invention
It is an object of the invention to provide handwritten signature location positioning side in a kind of bill based on depth convolutional neural networks
Method, can accurately calibrate the position containing handwritten signature in bill.
The technical solution adopted in the present invention is, handwritten signature location positioning in the bill based on depth convolutional neural networks
Method, specifically implements according to following steps：
Step 1：Build the platform based on the caffe deep learning framework comprising multiple convolutional neural networks models；
Step 2：Prepare the data set of bill；
Step 3：Training network obtains detection and localization model；
Step 4：The detection and localization model being obtained with the step 3 signature position hand-written to bill positioning to be detected.
The feature of the present invention also resides in：
Step 2 is specially：
Step 2.1：Bill is taken pictures, obtains the raw image data of bill, and sample is carried out to raw image data
This expansion；
Step 2.2：The all view data obtaining in step 2.1 are numbered and demarcate, calibrate in bill images
The coordinate of handwritten signature position, obtain the top left co-ordinate (Xmin, Ymin) of this signature position and bottom right angular coordinate (Xmax,
Ymax), and by the sequence number of all images and corresponding co-ordinate position information it is written in xml document；
Step 2.3：All view data are divided into training dataset and test data set, then training dataset is divided into instruction
Practice data and checking data.
In step 2.1, sample expansion is carried out to raw image data, including：
1. raw image data is carried out with the rotation of different angles and the translation of different directions；
2. linear interpolation scaling is carried out to raw image data；
3. raw image data is increased with salt-pepper noise, the Gaussian noise of varying strength.
Step 3 is specially：
Step 3.1：Image Adjusting in the data set obtaining in step 2 is become W1*H1The image of size simultaneously puts into ZF network
Front 5 layer by layer in do feature extraction, export 256The characteristic pattern of size；
Step 3.2：256 characteristic patterns being obtained with step 3.1 with the convolution kernel of 3*3 carry out convolution, obtain 256 dimensional features
Vector, as the ground floor of RPN；
Step 3.3：256 dimensional feature vectors that step 3.2 is obtained are input to classification layer and return two parallel-convolution of layer
Layer, according to the score height of prospect probability in positive sample, chooses high front 300 candidate frames of score；
Step 3.4：Using ROI_Pooling layer, 300 candidate frames in step 3.3 are mapped to the layer 5 of ZF network
On the 256 dimensional feature figures obtaining after convolution, obtain the characteristic pattern of size 6*6 after the normalizing of pond；
Step 3.5, the characteristic pattern of each 6*6 size is input to two continuously full articulamentum fc6, fc7, first passes through
Fc6 obtains 4096 dimensional features, then 4096 dimensional features are input in fc7, finally gives 1*4096 dimensional feature；
Step 3.6：1*4096 dimensional feature is input to two parallel full articulamentum cls_score layers, bbox_
Predict layer, cls_score layer is used for classifying, and exports the probability of background and the probability of K class sample, wherein, K is sample class
Number；Bbox_predict layer is used for adjusting candidate region location, and (x', y', w', the h') of output candidate frame, after x' is adjustment
Candidate frame upper left corner abscissa, y' is the candidate frame upper left corner vertical coordinate after adjustment, and w' is the width of the candidate frame after adjustment, h'
Height for the candidate frame after adjustment；
Step 3.7：Judge whether total iterationses are more than threshold value, if no more than threshold value, go to step 3.2；If
More than threshold value, then terminate.
Image Adjusting in data set in step 3.1 becomes W1*H1The image of size needs according to image size W* in data set
H calculating, that is,：
In step 3, during training network, the initial value of learning rate is set as lr=0.01, whenever current iteration number of times reaches
During the integral multiple of step value, learning rate is decayed once, when iterationses reach total iterationses, terminates；Lr after decay
=lr*gamma, wherein, gamma=0.1, iterationses≤total iterationses.
In step 3, during training network, mini-batch size is set to 256.
The invention has the beneficial effects as follows：Handwritten signature location positioning in the bill based on depth convolutional neural networks for the present invention
Method, uses deep neural network to carry out framing, more equal in the speed and accuracy of positioning than traditional method
It is improved, accuracy can reach 90.9%, speed can accomplish real-time positioning substantially, positioning one pictures need
0.3s；And the bill images data base of the present invention contains various data samples, and data sample has multiformity, makes
The accuracy that must position has risen.
Brief description
Fig. 1 is the normal bill images of collection in the inventive method；
Fig. 2 is 45 ° of bill images of rotation of collection in the inventive method；
Fig. 3 is collection plus salt-pepper noise bill images in the inventive method；
Fig. 4 is the bill images to be detected rotating 90 ° in the inventive method；
Fig. 5 is the testing result figure of Fig. 4；
Fig. 6 is the bill images to be detected of 180 ° of rotation in the inventive method；
Fig. 7 is the testing result figure of Fig. 6；
Fig. 8 is the bill images to be detected of 45 ° of rotation in the inventive method；
Fig. 9 is the testing result figure of Fig. 8；
Figure 10 is in the inventive method plus the bill images to be detected of Gaussian noise；
Figure 11 is the testing result figure of Figure 10；
Figure 12 is in the inventive method plus the bill images to be detected of salt-pepper noise；
Figure 13 is the testing result figure of Figure 12.
Specific embodiment
The present invention is described in detail with reference to the accompanying drawings and detailed description.
Handwritten signature location positioning method in the bill based on depth convolutional neural networks for the present invention, specifically according to following step
Rapid enforcement：
Step 1：Build to be based under ubuntu system environmentss or Windos environment and comprise multiple convolutional neural networks moulds
The platform of the caffe deep learning framework of type；
Step 2：Prepare the data set of bill, specially：
Step 2.1, due to being all to be taken pictures on mobile phone or camera upload by user oneself in the image in true environment
To billing system, so when preparing pictures, needing the photo that the mobile phone considering different resolution shoots, and shoot
The condition such as ambient lighting.The present invention is shot to bill using the mobile phone of multiple different resolutions, and this part mobile phone is clapped
The image taken the photograph is referred to as raw image data.In order that view data is abundant enough and can meet various practical situations, the present invention
Sample expansion has been carried out to raw image data：1. raw image data is carried out with the rotation of different angles and the flat of different directions
Move；2. linear interpolation scaling is carried out to raw image data, allow for the image that different cameral collects of different sizes；3. right
Raw image data increases the salt-pepper noise of varying strength, Gaussian noise；
As Fig. 1-3 provides the part bill sample image collecting, including the bill of normal bill images, rotation
As, contain noisy bill images.
Step 2.2：The all view data obtaining in step 2.1 are numbered and demarcate, calibrate in bill images
The coordinate of handwritten signature position, obtain the top left co-ordinate (Xmin, Ymin) of this signature position and bottom right angular coordinate (Xmax,
Ymax), and by all picture numbers and corresponding co-ordinate position information it is written in xml document；
Step 2.3：Training dataset trainval and test data set are randomly divided into all ready view data
Two partial datas of test, are provided with trainval data set and account for 8/10, the test data set of whole data set accounting for and entirely counting
According to collection 2/10.It is divided into train data and val data in trainval data set, wherein trian data is used to instruct again
Practice, 4/5, the val data accounting for trainval data set is used to do and verifies, accounts for the 1/5 of trainval data set.
Step 3：Training network obtains detection and localization model and training parameter is optimized
Step 3.1：Image Adjusting in the data set obtaining in step 2 is become W1*H1(being 600*800 in the present invention) is big
Little image and before putting into ZF network 5 layer by layer in do feature extraction, export 256(being 37*50 in the present invention) size
Characteristic pattern.In data set, the size of Image Adjusting needs length-width ratio (the image longest edge/figure according to image in data images
As minor face) calculating, if input picture size is W*H, the image size after adjustment is W1*H1, relational expression is：
The reason in data set, image size is adjusted to 600*800：
Image in data set is not of uniform size, when calculating image size after normalizing, selectes that to account for whole data set number more
Several types size image, and in data set the ratio of image longest edge and image minor face to account for whole data set more
Several types image, qualified inclusion 600*800,1200*1600,1500*2000,2000*2600,3000*
4000, using these as adjustment after image size, the image after adjustment need to carry out convolutional calculation it is contemplated that amount of calculation big
Little, GPU memory size, need to choose less adjustment ratio as far as possible, be tentatively revised as 600*800,1200*1600 this two big
Little image is trained, and the image choosing 600*800 size is trained accuracy for 90.84%, chooses 1200*1600 big
The accuracy that little image is trained is 90.86%, and accuracy the latter only improves 0.02% than the former, but the training time with
And the latter is more more complex than the former on computation complexity, therefore finally choose this ratio of 600*800 as the ratio of adjustment.
Step 3.2：Carry out convolution with 256 characteristic patterns that the convolution kernel (sliding window) of 3*3 is obtained with step 3.1, because
For on the region of this 3*3, each characteristic pattern obtaining 1 dimensional vector, 256 characteristic patterns can get 256 Wei Te
Levy vector, as the ground floor of RPN (region proposal network, region candidate network)；
3*3 sliding window center position, 3 kinds of yardsticks (128,256,512) of corresponding prediction input picture, 3 kinds of length-width ratios (1：
2、2：1、7：1) target area candidate frame, the mechanism of this mapping is referred to as anchor point, creates k=9 anchor point.It is each 3*3
Region can produce 9 target area candidate frames.So for the characteristic pattern of the 37*50 in the present invention, a total of about 20000
(37*50*9) individual anchor point, that is, predict 20000 target area candidate frames over an input image.
The yardstick of target area candidate frame has three (128,256,512), and this essence refers to that the area of candidate frame is
128*128,256*256,512*512, the size after Image Adjusting normalization in this candidate frame size and step 3.1 has
Target in image is enclosed in interior by pass, the as far as possible area of candidate frame.
The length-width ratio of target area candidate frame selects 1：2、2：1 and 7：1 these three ratios are according to each figure in data set
Width (W as interior target candidate framebox) and high (Hbox) ratio draw, choose three most ratios of wide high proportion picture number and make
The ratio of width to height for candidate frameWherein, Wbox=Xmax-Xmin, Hbox=Ymax-Ymin, (Xmin, Ymin), (Xmax,
Ymax) it is the upper left corner of handwritten signature position demarcated in step 2.2 and lower right corner coordinate figure.
Step 3.3：256 dimensional feature vectors are input to two parallel-convolution layers, that is, classification layer and recurrence layer, are respectively used to
Classification and frame return.For local, this two-layer is fully-connected network；For the overall situation, due to network in all positions (altogether
37*50) parameter identical, so the actual convolutional network with a size of 1 × 1 is realized.It should be noted that：Not explicit
Ground extracts any candidate window, completely completes to judge using network itself and revises.
To each candidate frame, layer of classifying exports, from 256 dimensional features, the probability belonging to foreground and background, and to each candidate
Frame is demarcated：Positive sample be overlapping with real estate be more than 0.7, negative sample be overlapping with real estate be less than 0.3, reservation just
Sample；
Return layer simultaneously and export 4 translations zooming parameter (x, y, w, h) from 256 dimensional features, wherein, x is that candidate frame is left
Upper angle abscissa, y is candidate frame upper left corner vertical coordinate, and w is the width of candidate frame, and h is the height of candidate frame, this four coordinate units
Element is used for determining target location.
Through step 3.2, the present invention predicts over an input image and obtains 20000 candidate frames, after step 3.3,
20000 predicting candidate frames are left 2000 about candidate frames, finally according to the score height of prospect probability in positive sample, choose
High front 300 candidate frames of score.
Step 3.4：Using ROI_Pooling layer, 300 candidate frames in step 3.3 are mapped to the layer 5 of ZF network
On the 256 dimensional feature figures obtaining after convolution, obtain the characteristic pattern of size 6*6 after the normalizing of pond.
ROI_Pooling layer is exactly to realize from artwork area maps to conv5 region after-bay to the work(of fixed size
Energy.
Calculate the coordinate that the candidate frame of prediction is mapped on characteristic pattern first, that is, original coordinates are multiplied by 1/16th, so
It is directed to each afterwards to export being calculated, the 300 different size of candidate frames that will map on characteristic pattern carry out pond
Change, then the result unification of Chi Huahou is normalized to the characteristic pattern of size 6*6.
Step 3.5, the characteristic pattern (300 characteristic patterns altogether) of each 6*6 size is input to two continuously entirely connects
Stratum fc6, fc7, this two full articulamentums are not to be continuously parallel, first pass through fc6 and obtain 4096 dimensional features, then by 4096
Dimensional feature is input in fc7, finally gives 1*4096 dimensional feature.
Step 3.6：1*4096 dimensional feature is input to two parallel full articulamentum cls_score layers, bbox_
Predict layer, cls_score layer is used for classifying, and exports the probability of background and the probability of K class sample, wherein, K is sample class
Number；Bbox_predict layer is used for adjusting candidate region location, and (x', y', w', the h') of output candidate frame, after x' is adjustment
Candidate frame upper left corner abscissa, y' is the candidate frame upper left corner vertical coordinate after adjustment, and w' is the width of the candidate frame after adjustment, h'
Height for the candidate frame after adjustment；
Step 3.7：Judge whether total iterationses (for 8000 in the present invention) are more than threshold value, if no more than threshold value,
Go to step 3.2；If greater than threshold value, then terminate.
The selection of total iterationses：Observe the value of loss in the training process, when the value of loss no longer significantly declines, becomes
In stablizing, we may be selected by current iterationses iterationses the most final.
In training network, the initial value of learning rate is set as lr=0.01, whenever current iteration number of times reaches step value
During the integral multiple of (present invention is 6000), learning rate is decayed once, when iterationses reach total iterationses, terminates；Decline
Lr=lr*gamma after subtracting, wherein, gamma=0.1, iterationses≤total iterationses.
When being optimized with gradient descent algorithm, in the more new regulation of weight, a coefficient can be multiplied by before gradient terms,
This coefficient is just learning rate lr.If learning rate is too little, network convergence can be made excessively slow, if learning rate is too big,
Cost function then can be led to vibrate, a reasonable strategy is first learning rate to be set to 0.01, Ran Houguan in practice
Examine the trend of training cost, if training cost is reducing, that can tune up learning rate step by step, such as 0.1,
1.0….If training cost is increasing, that is reduced by learning rate, such as 0.001,0.0001 ....Through said method
Determine the value of learning rate.
It is relevant with step-length when learning rate decays, and how much relevant with gamma reduces.When selecting step-length, Ke Yijin
Possible close total iterationses.
During training network, mini-batch size is set to 256, using the more new regulation of weight during mini-batch is：
Namely the gradient of 256 samples is averaged.
When using mini-batch, all samples in a batch can be placed in a matrix for we, utilize
The calculating of accelerating gradient is carried out in linear algebra storehouse, and this is one of Project Realization optimization method.
One big batch size, can make full use of matrix, linear algebra storehouse come the acceleration being calculated, batch
Size is less, then acceleration effect may be more inconspicuous.Batch size is not the bigger the better, too big, and the renewal of weight will
Less frequent, lead to optimization process too very long.General batch size size is 256, if image data collection less, GPU
Below internal memory 4G it may be considered that change little by batch size.
Step 4：The detection and localization model being obtained with the step 3 signature position hand-written to bill positioning to be detected, that is, wrap
Include top left co-ordinate and the bottom right angular coordinate of the rectangle frame of handwritten signature position.
Fig. 4-Figure 13 is the bill images to be detected and use basis with different rotary angle and different noise types
Handwritten signature position at inventive method positioning uses rectangle frame labelling figure in the figure.As can be seen that the robust of the inventive method
Property is very good, can overcome the impact to positioning result for situations such as rotation, noise, has the fast spy of accurate positioning, locating speed
Point.
Claims (7)
1. in the bill based on depth convolutional neural networks handwritten signature location positioning method it is characterised in that specifically according to
Lower step is implemented：
Step 1：Build the platform based on the caffe deep learning framework comprising multiple convolutional neural networks models；
Step 2：Prepare the data set of bill；
Step 3：Training network obtains detection and localization model；
Step 4：The detection and localization model being obtained with the step 3 signature position hand-written to bill positioning to be detected.
2. handwritten signature location positioning method in the bill based on depth convolutional neural networks according to claim 1, its
It is characterised by, described step 2 is specially：
Step 2.1：Bill is taken pictures, obtains the raw image data of bill, and sample expansion is carried out to raw image data
Fill；
Step 2.2：The all view data obtaining in step 2.1 are numbered and demarcate, calibrate hand-written in bill images
The coordinate of signature position, obtains top left co-ordinate (Xmin, Ymin) and the bottom right angular coordinate (Xmax, Ymax) of this signature position,
And the sequence number of all images and corresponding co-ordinate position information are written in xml document；
Step 2.3：All view data are divided into training dataset and test data set, then training dataset is divided into training number
According to checking data.
3. handwritten signature location positioning method in the bill based on depth convolutional neural networks according to claim 2, its
It is characterised by, in described step 2.1, sample expansion is carried out to raw image data, including：
1. raw image data is carried out with the rotation of different angles and the translation of different directions；
2. linear interpolation scaling is carried out to raw image data；
3. raw image data is increased with salt-pepper noise, the Gaussian noise of varying strength.
4. handwritten signature location positioning method in the bill based on depth convolutional neural networks according to claim 1, its
It is characterised by, described step 3 is specially：
Step 3.1：Image Adjusting in the data set obtaining in step 2 is become W1*H1The image of size is simultaneously put into 5 before ZF network
Do feature extraction in layer by layer, export 256The characteristic pattern of size；
Step 3.2：256 characteristic patterns being obtained with step 3.1 with the convolution kernel of 3*3 carry out convolution, obtain 256 dimensional features to
Amount, as the ground floor of RPN；
Step 3.3：256 dimensional feature vectors that step 3.2 is obtained are input to classification layer and return two parallel-convolution layer of layer, root
According to the score height of prospect probability in positive sample, choose high front 300 candidate frames of score；
Step 3.4：Using ROI_Pooling layer, 300 candidate frames in step 3.3 are mapped to the layer 5 convolution of ZF network
On the 256 dimensional feature figures obtaining afterwards, obtain the characteristic pattern of size 6*6 after the normalizing of pond；
Step 3.5, the characteristic pattern of each 6*6 size is input to two continuously full articulamentum fc6, fc7, first passes through fc6
Obtain 4096 dimensional features, then 4096 dimensional features are input in fc7, finally give 1*4096 dimensional feature；
Step 3.6：1*4096 dimensional feature is input to two parallel full articulamentum cls_score layers, bbox_predict layers,
Cls_score layer is used for classifying, and exports the probability of background and the probability of K class sample, wherein, K is sample class number；bbox_
Predict layer is used for adjusting candidate region location, (x', y', w', the h') of output candidate frame, and x' is that the candidate frame after adjustment is left
Upper angle abscissa, y' is the candidate frame upper left corner vertical coordinate after adjustment, and w' is the width of the candidate frame after adjustment, after h' is adjustment
Candidate frame height；
Step 3.7：Judge whether total iterationses are more than threshold value, if no more than threshold value, go to step 3.2；If greater than
Threshold value, then terminate.
5. handwritten signature location positioning method in the bill based on depth convolutional neural networks according to claim 4, its
It is characterised by, the Image Adjusting in data set in described step 3.1 becomes W1*H1The image of size needs according to image in data set
Size W*H calculating, that is,：
6. handwritten signature location positioning method in the bill based on depth convolutional neural networks according to claim 4, its
It is characterised by, in described step 3, during training network, the initial value of learning rate is set as lr=0.01, whenever current iteration number of times
Reach step value integral multiple when, learning rate decay once, when iterationses reach total iterationses, terminate；After decay
Lr=lr*gamma, wherein, gamma=0.1, iterationses≤total iterationses.
7. handwritten signature location positioning method in the bill based on depth convolutional neural networks according to claim 4, its
It is characterised by, in described step 3, during training network, mini-batch size is set to 256.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610841643.8A CN106469304A (en) | 2016-09-22 | 2016-09-22 | Handwritten signature location positioning method in bill based on depth convolutional neural networks |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610841643.8A CN106469304A (en) | 2016-09-22 | 2016-09-22 | Handwritten signature location positioning method in bill based on depth convolutional neural networks |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN106469304A true CN106469304A (en) | 2017-03-01 |
Family
ID=58230165
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610841643.8A Pending CN106469304A (en) | 2016-09-22 | 2016-09-22 | Handwritten signature location positioning method in bill based on depth convolutional neural networks |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106469304A (en) |
Cited By (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106960186A (en) * | 2017-03-17 | 2017-07-18 | 王宇宁 | Ammunition recognition methods based on depth convolutional neural networks |
| CN107229929A (en) * | 2017-04-12 | 2017-10-03 | 西安电子科技大学 | A kind of license plate locating method based on R CNN |
| CN107368845A (en) * | 2017-06-15 | 2017-11-21 | 华南理工大学 | A kind of Faster R CNN object detection methods based on optimization candidate region |
| CN107578439A (en) * | 2017-07-19 | 2018-01-12 | 阿里巴巴集团控股有限公司 | Generate the method, apparatus and equipment of target image |
| CN107798301A (en) * | 2017-10-13 | 2018-03-13 | 上海眼控科技股份有限公司 | A kind of signature detection system and method for vehicle annual test |
| CN107977665A (en) * | 2017-12-15 | 2018-05-01 | 北京科摩仕捷科技有限公司 | The recognition methods of key message and computing device in a kind of invoice |
| CN108564035A (en) * | 2018-04-13 | 2018-09-21 | 杭州睿琪软件有限公司 | The method and system for the information recorded on identification document |
| CN108898188A (en) * | 2018-07-06 | 2018-11-27 | 四川奇迹云科技有限公司 | A kind of image data set aid mark system and method |
| CN108921057A (en) * | 2018-06-19 | 2018-11-30 | 厦门大学 | Prawn method for measuring shape of palaemon, medium, terminal device and device based on convolutional neural networks |
| CN109813276A (en) * | 2018-12-19 | 2019-05-28 | 五邑大学 | A kind of antenna for base station has a down dip angle measuring method and its system |
| CN111160142A (en) * | 2019-12-14 | 2020-05-15 | 上海交通大学 | Certificate bill positioning detection method based on numerical prediction regression model |
| WO2020133966A1 (en) * | 2018-12-29 | 2020-07-02 | 北京市商汤科技开发有限公司 | Anchor determining method and apparatus, and electronic device and storage medium |
| CN111488964A (en) * | 2019-01-29 | 2020-08-04 | 北京市商汤科技开发有限公司 | Image processing method and device and neural network training method and device |
| CN112102281A (en) * | 2020-09-11 | 2020-12-18 | 哈尔滨市科佳通用机电股份有限公司 | Truck brake cylinder fault detection method based on improved Faster Rcnn |
| TWI777188B (en) * | 2020-07-07 | 2022-09-11 | 新光人壽保險股份有限公司 | Contract signature authentication method and device |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104298976A (en) * | 2014-10-16 | 2015-01-21 | 电子科技大学 | License plate detection method based on convolutional neural network |
| CN104966097A (en) * | 2015-06-12 | 2015-10-07 | 成都数联铭品科技有限公司 | Complex character recognition method based on deep learning |
| CN105488468A (en) * | 2015-11-26 | 2016-04-13 | 浙江宇视科技有限公司 | Method and device for positioning target area |
| CN105550701A (en) * | 2015-12-09 | 2016-05-04 | 福州华鹰重工机械有限公司 | Real-time image extraction and recognition method and device |
| US9400919B2 (en) * | 2014-05-27 | 2016-07-26 | Beijing Kuangshi Technology Co., Ltd. | Learning deep face representation |
| CN105868774A (en) * | 2016-03-24 | 2016-08-17 | 西安电子科技大学 | Selective search and convolutional neural network based vehicle logo recognition method |
| CN105893952A (en) * | 2015-12-03 | 2016-08-24 | 无锡度维智慧城市科技股份有限公司 | Hand-written signature identifying method based on PCA method |
-
2016
- 2016-09-22 CN CN201610841643.8A patent/CN106469304A/en active Pending
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US9400919B2 (en) * | 2014-05-27 | 2016-07-26 | Beijing Kuangshi Technology Co., Ltd. | Learning deep face representation |
| CN104298976A (en) * | 2014-10-16 | 2015-01-21 | 电子科技大学 | License plate detection method based on convolutional neural network |
| CN104966097A (en) * | 2015-06-12 | 2015-10-07 | 成都数联铭品科技有限公司 | Complex character recognition method based on deep learning |
| CN105488468A (en) * | 2015-11-26 | 2016-04-13 | 浙江宇视科技有限公司 | Method and device for positioning target area |
| CN105893952A (en) * | 2015-12-03 | 2016-08-24 | 无锡度维智慧城市科技股份有限公司 | Hand-written signature identifying method based on PCA method |
| CN105550701A (en) * | 2015-12-09 | 2016-05-04 | 福州华鹰重工机械有限公司 | Real-time image extraction and recognition method and device |
| CN105868774A (en) * | 2016-03-24 | 2016-08-17 | 西安电子科技大学 | Selective search and convolutional neural network based vehicle logo recognition method |
Non-Patent Citations (8)
| Title |
|---|
| LUIZ G. HAFEMANN等: "Writer-independent feature learning for Offline Signature Verification using Deep Convolutional Neural Networks", 《2016 INTERNATIONAL JOINT CONFERENCE ON NEURAL NETWORKS (IJCNN)》 * |
| MATTHEW D. ZEILER等: "Visualizing and Understanding Convolutional Networks", 《ECCV 2014》 * |
| ROSS GIRSHICK: "Fast R-CNN", 《ICCV 2015》 * |
| ROSS GIRSHICK等: "Region-Based Convolutional Networks for Accurate Object Detection and Segmentation", 《IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE》 * |
| SHAOQING REN等: "Faster R-CNN: towards real-time object detection with region proposal networks", 《NIPS"15》 * |
| 何柳: "表单识别中的关键问题研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
| 杨名等: "甚高速区域卷积神经网络的船舶视频目标识别算法", 《2016年全国通信软件学术会议程序册与交流文集》 * |
| 陈致远: "基于区域梯度统计分析与卷积神经网络的条码定位算法研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (23)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106960186A (en) * | 2017-03-17 | 2017-07-18 | 王宇宁 | Ammunition recognition methods based on depth convolutional neural networks |
| CN107229929A (en) * | 2017-04-12 | 2017-10-03 | 西安电子科技大学 | A kind of license plate locating method based on R CNN |
| CN107368845A (en) * | 2017-06-15 | 2017-11-21 | 华南理工大学 | A kind of Faster R CNN object detection methods based on optimization candidate region |
| CN107578439B (en) * | 2017-07-19 | 2020-04-28 | 创新先进技术有限公司 | Method, device and equipment for generating target image |
| CN107578439A (en) * | 2017-07-19 | 2018-01-12 | 阿里巴巴集团控股有限公司 | Generate the method, apparatus and equipment of target image |
| CN107798301A (en) * | 2017-10-13 | 2018-03-13 | 上海眼控科技股份有限公司 | A kind of signature detection system and method for vehicle annual test |
| CN107977665A (en) * | 2017-12-15 | 2018-05-01 | 北京科摩仕捷科技有限公司 | The recognition methods of key message and computing device in a kind of invoice |
| CN108564035A (en) * | 2018-04-13 | 2018-09-21 | 杭州睿琪软件有限公司 | The method and system for the information recorded on identification document |
| CN108564035B (en) * | 2018-04-13 | 2020-09-25 | 杭州睿琪软件有限公司 | Method and system for identifying information recorded on document |
| US10977513B2 (en) | 2018-04-13 | 2021-04-13 | Hangzhou Glorify Software Limited | Method, system and computer readable storage medium for identifying information carried on sheet |
| CN108921057A (en) * | 2018-06-19 | 2018-11-30 | 厦门大学 | Prawn method for measuring shape of palaemon, medium, terminal device and device based on convolutional neural networks |
| CN108921057B (en) * | 2018-06-19 | 2021-06-01 | 厦门大学 | Convolutional neural network-based prawn form measuring method, medium, terminal equipment and device |
| CN108898188A (en) * | 2018-07-06 | 2018-11-27 | 四川奇迹云科技有限公司 | A kind of image data set aid mark system and method |
| CN109813276A (en) * | 2018-12-19 | 2019-05-28 | 五邑大学 | A kind of antenna for base station has a down dip angle measuring method and its system |
| US11301726B2 (en) | 2018-12-29 | 2022-04-12 | Beijing Sensetime Technology Development Co., Ltd. | Anchor determination method and apparatus, electronic device, and storage medium |
| WO2020133966A1 (en) * | 2018-12-29 | 2020-07-02 | 北京市商汤科技开发有限公司 | Anchor determining method and apparatus, and electronic device and storage medium |
| CN111488964A (en) * | 2019-01-29 | 2020-08-04 | 北京市商汤科技开发有限公司 | Image processing method and device and neural network training method and device |
| CN111488964B (en) * | 2019-01-29 | 2024-07-16 | 北京市商汤科技开发有限公司 | Image processing method and device, and neural network training method and device |
| CN111160142A (en) * | 2019-12-14 | 2020-05-15 | 上海交通大学 | Certificate bill positioning detection method based on numerical prediction regression model |
| CN111160142B (en) * | 2019-12-14 | 2023-07-11 | 上海交通大学 | Certificate bill positioning detection method based on numerical prediction regression model |
| TWI777188B (en) * | 2020-07-07 | 2022-09-11 | 新光人壽保險股份有限公司 | Contract signature authentication method and device |
| CN112102281B (en) * | 2020-09-11 | 2021-07-06 | 哈尔滨市科佳通用机电股份有限公司 | Truck brake cylinder fault detection method based on improved Faster Rcnn |
| CN112102281A (en) * | 2020-09-11 | 2020-12-18 | 哈尔滨市科佳通用机电股份有限公司 | Truck brake cylinder fault detection method based on improved Faster Rcnn |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106469304A (en) | Handwritten signature location positioning method in bill based on depth convolutional neural networks | |
| CN111126472B (en) | SSD (solid State disk) -based improved target detection method | |
| CN110298387A (en) | Incorporate the deep neural network object detection method of Pixel-level attention mechanism | |
| CN113160062B (en) | Infrared image target detection method, device, equipment and storage medium | |
| CN107871134A (en) | A kind of method for detecting human face and device | |
| CN110298298A (en) | Target detection and the training method of target detection network, device and equipment | |
| CN106228125B (en) | Method for detecting lane lines based on integrated study cascade classifier | |
| CN111368769B (en) | Ship multi-target detection method based on improved anchor point frame generation model | |
| CN109447034A (en) | Traffic mark detection method in automatic Pilot based on YOLOv3 network | |
| CN110119728A (en) | Remote sensing images cloud detection method of optic based on Multiscale Fusion semantic segmentation network | |
| CN110443208A (en) | YOLOv 2-based vehicle target detection method, system and equipment | |
| CN110276269A (en) | A kind of Remote Sensing Target detection method based on attention mechanism | |
| CN109711474A (en) | A kind of aluminium material surface defects detection algorithm based on deep learning | |
| CN110163187A (en) | Remote road traffic sign detection recognition methods based on F-RCNN | |
| CN109800628A (en) | A kind of network structure and detection method for reinforcing SSD Small object pedestrian detection performance | |
| CN113486764B (en) | Pothole detection method based on improved YOLOv3 | |
| CN108334847A (en) | A kind of face identification method based on deep learning under real scene | |
| CN109800629A (en) | A kind of Remote Sensing Target detection method based on convolutional neural networks | |
| CN110135430A (en) | A kind of aluminium mold ID automatic recognition system based on deep neural network | |
| CN110298265A (en) | Specific objective detection method in a kind of elevator based on YOLO neural network | |
| CN109598234A (en) | Critical point detection method and apparatus | |
| CN106250895A (en) | A kind of remote sensing image region of interest area detecting method | |
| CN109711416A (en) | Target identification method, device, computer equipment and storage medium | |
| CN109087337B (en) | Long-time target tracking method and system based on hierarchical convolution characteristics | |
| CN108122003A (en) | A kind of Weak target recognition methods based on deep neural network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20170301 |