CN106407959A - Low-illumination complicated background license plate positioning method based on wavelet transform and SVM - Google Patents

Low-illumination complicated background license plate positioning method based on wavelet transform and SVM Download PDF

Info

Publication number
CN106407959A
CN106407959A CN201610976425.5A CN201610976425A CN106407959A CN 106407959 A CN106407959 A CN 106407959A CN 201610976425 A CN201610976425 A CN 201610976425A CN 106407959 A CN106407959 A CN 106407959A
Authority
CN
China
Prior art keywords
image
license plate
wavelet
svm
area
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
Application number
CN201610976425.5A
Other languages
Chinese (zh)
Inventor
陈蓉
姜德财
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Hunan Yuan Xin Electro-Optical Technology Inc (us) 62 Martin Road Concord Massachusetts 017
Original Assignee
Hunan Yuan Xin Electro-Optical Technology Inc (us) 62 Martin Road Concord Massachusetts 017
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Hunan Yuan Xin Electro-Optical Technology Inc (us) 62 Martin Road Concord Massachusetts 017 filed Critical Hunan Yuan Xin Electro-Optical Technology Inc (us) 62 Martin Road Concord Massachusetts 017
Priority to CN201610976425.5A priority Critical patent/CN106407959A/en
Publication of CN106407959A publication Critical patent/CN106407959A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/50—Context or environment of the image
    • G06V20/52—Surveillance or monitoring of activities, e.g. for recognising suspicious objects
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing
    • G06V10/25—Determination of region of interest [ROI] or a volume of interest [VOI]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/42—Global feature extraction by analysis of the whole pattern, e.g. using frequency domain transformations or autocorrelation
    • G06V10/431—Frequency domain transformation; Autocorrelation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
    • G06V10/443—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components by matching or filtering
    • G06V10/446—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components by matching or filtering using Haar-like filters, e.g. using integral image techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/60—Type of objects
    • G06V20/62—Text, e.g. of license plates, overlay texts or captions on TV images
    • G06V20/625—License plates

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Data Mining & Analysis (AREA)
  • Multimedia (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Image Analysis (AREA)

Abstract

The invention provides a low-illumination complicated background license plate positioning method based on wavelet transform and SVM, which relates to the field of computer vision. The low-illumination complicated background license plate positioning method comprises the steps of: utilizing wavelet transform to perform multi-level wavelet decomposition on an image to obtain scale coefficients and wavelet coefficients, adopting an MSR method for processing the scale coefficients to adjust the brightness of the low-illumination image, and adopting a threshold enhancement method for processing the wavelet coefficients of layers to enhance the image, thereby acquiring the clear image with rich detail information and high contrast and being conductive to perform license plate positioning accurately; utilizing a vertical edge detection Bernsen operator for carrying out edge detection, combining with a vertical projection method for carrying out rough positioning on the license plate quickly to obtain the width of a license plate region, and filtering out most of the irrelevant background region; and finally extracting spatial domain features and discrete cosine domain features of the image, utilizing an SVM algorithm for classification training so as to position the license plate region, and removing overlapping frames between the license plate region to obtain the final precise positioning effect.

Description

Wavelet transformation and SVM (support vector machine) -based low-illumination complex background license plate positioning method
Technical Field
The invention relates to the field of computer vision, in particular to a low-illumination complex background license plate positioning method based on wavelet transformation and SVM.
Background
An intelligent traffic system has become a mainstream direction of traffic management development, and has been taken as a License Plate Recognition technology (LPR) for realizing traffic management automation, and the LPR system is widely applied to occasions such as electronic toll stations and parking lot vehicle management. The license plate recognition system mainly comprises the following components of license plate image acquisition and preprocessing, license plate region positioning and extraction, license plate character segmentation and recognition and the like.
The license plate positioning effect can directly influence the following character segmentation and character recognition. The commonly used license plate positioning methods include a license plate positioning method based on edge features, a license plate positioning method based on color features, a license plate positioning method based on texture features and the like. Edge-based positioning methods are generally fast and simple, but are easily interfered by the background, and often have poor positioning effects when the background has more complex interference edges. The license plate positioning method based on the color characteristics is accurate in positioning and high in precision. However, when the color of the vehicle body is closer to the background color, the license plate positioning fails, so that the method has obvious limitation. The positioning method based on the texture features can overcome the problems of edge blurring, color distortion and the like, but the calculation amount is large. Because the license plate recognition system is often applied to an outdoor environment, license plate positioning is affected by many factors in the actual application process, such as fuzzy license plates, nonuniform hanging positions of the license plates, nonuniform and insufficient illumination, unfixed size of the license plates, complex background and the like, which bring certain difficulties to the license plate positioning.
Aiming at the difficulty and inaccuracy of license plate positioning caused by the loss of license plate information under low illumination and excessive interference information in a complex background, a wavelet transformation and SVM-based license plate positioning method under low illumination with the complex background is provided. The method comprises the steps of firstly, utilizing wavelet transformation to carry out multi-level wavelet decomposition on an image to obtain a scale coefficient and a wavelet coefficient, adopting an MSR (minimum shift register) method to process the scale coefficient to adjust the brightness of a low-illumination image, and adopting a threshold enhancement method to process the wavelet coefficient of each layer to enhance the image, so that a clear image with rich detail information and high contrast is obtained, and license plate positioning is facilitated accurately; then, edge detection is carried out by utilizing a vertical edge detection Bernsen operator, coarse positioning of the license plate is rapidly carried out by combining a vertical projection method, the width of the license plate region is obtained, most irrelevant background regions are filtered, and the speed of accurate positioning is favorably improved; and finally, extracting the spatial domain characteristics and the discrete cosine domain characteristics of the image, performing classification training by using an SVM algorithm, thereby positioning a license plate region, and then removing the overlapping part of rectangular frames of the license plate region to obtain the final accurate positioning effect.
Disclosure of Invention
The technical problem to be solved by the invention is as follows: aiming at the specific problem of license plate positioning, a license plate positioning method based on wavelet and SVM with low illumination and complex background is provided.
In order to solve the technical problems, the invention adopts the technical scheme that: a license plate positioning method based on wavelet transformation and SVM low-illumination complex background is characterized by comprising the following steps:
step 1) using wavelet transformation to perform image enhancement on a low-illumination image to obtain an enhanced image with abundant detail information and high contrast, and adopting different threshold enhancement algorithms for wavelet coefficients of different levels, so that the detail characteristics of the image are enhanced and noise is suppressed, and the method comprises the following steps:
1.1, performing multi-level wavelet decomposition on the image by utilizing wavelet transformation to obtain a scale coefficient phi (t) and a wavelet coefficient psi (t);
1.2, processing the scale coefficient by adopting an MSR method to adjust the brightness of the image;
1.3, processing wavelet coefficients of each layer by adopting a threshold enhancement method to enhance the detail characteristics of the image;
step 2) carrying out binarization processing on the enhanced image, carrying out edge detection by using a vertical edge detection Bernsen operator, and quickly carrying out coarse positioning on the license plate by combining a vertical projection method to obtain the width of the license plate region, thereby filtering out most irrelevant background regions, being beneficial to quickly realizing SVM classification training and improving the speed of the system;
step 3) extracting the airspace characteristics and discrete cosine domain characteristics of the license plate region obtained in the step 2), and performing classification training by using an SVM algorithm so as to locate the license plate region, wherein the method comprises the following steps:
3.1, extracting the spatial domain characteristic and the discrete cosine characteristic of the image to be used as a characteristic vector of SVM training;
3.2, carrying out classification training by utilizing an SVM algorithm, and positioning a license plate area;
and 4) detecting the overlapped rectangular frames of the license plate area, and accurately positioning the license plate area.
As a further improvement of the present invention, in step 1), the method for adjusting the brightness of the image by processing the scale coefficient by using the MSR method is as follows:
in the above formula, R (x, y) is the output of the MSR method; i (x, y) is image data;is the convolution operator; w is aeAs a weight, ∑ w is satisfiede1, e denotes the order, i.e. the number of weights, F (x, y) K x exp (- (x)2+y2)/ce 2) K satisfies: (x, y) dxdy ═ 1; c. CeDetermining the shape of the Gaussian curved surface; e is the number of Gaussian curved surfaces with different shapes, namely the number of scales, and different E correspond to different ce。
As a further improvement of the present invention, in step 1), the processing wavelet coefficients of each layer by using a threshold enhancement method to enhance the detail features of the image includes:
selecting 3-time B-spline wavelet with the length of 5 to carry out four-level wavelet decomposition on the image, and respectively transforming wavelet components with different scales through different threshold algorithms according to different characteristics of wavelet coefficients of each layer;
within the first-level wavelet coefficient, image detail features are enhanced by adopting a dual threshold:
in the above formula, T1And T2Is the threshold, T1＜T2，G＝8，WinAnd WoutIs the wavelet coefficient before and after transformation;
in the second and third wavelet coefficients, adopting an adaptive threshold enhancement algorithm:
in the formula, T3And T4Is the threshold, T4＜T3，G＝10，WinAnd WoutIs the wavelet coefficient before and after transformation;
and (3) in a fourth-level wavelet coefficient, processing by adopting a single threshold algorithm:
in the formula, T5Is the threshold, G ═ 8, WinAnd WoutAre wavelet coefficients before and after the transform.
As a further improvement of the present invention, in the step 2), the lower 2/3 part of the license plate image is vertically projected, the license plate is scanned from bottom to top, and the width of the license plate position is roughly located by combining the vertical projection method, and the step includes:
(1) vertically projecting the processed binary image, scanning the vertically projected license plate image, counting the absolute difference of the number of black pixel points of two adjacent columns in the projected image, and simultaneously recording the coordinates of the previous column corresponding to each absolute difference;
(2) and arranging the absolute differences by using a quick ordering method, ordering the corresponding coordinates, taking 18 coordinate values with the maximum absolute differences, ordering from small to large, removing abrupt points by using a least square method, and obtaining the width of the license plate by using the remaining area corresponding to the coordinate values, namely the area of the license plate in the vertical direction, namely the area of the vertical strip.
As a further improvement of the present invention, in the step 3), a gray level co-occurrence matrix is used for extracting image surface texture feature parameters for spatial domain features, and the gray level values of the image f (i, j) are gradedWith the highest gray level of NgThe width of the image in the horizontal direction is Lx＝{1,2,...,NxLength of image in vertical direction is Ly＝{1,2,...,NyAnd recording a gray level set of the image as G ═ 1,2g}. The image f (i, j) to be analyzed can be considered as Lx×LyOne transformation to G, i.e. each point L in the image f (i, j)x×LyCorresponding to a gray value belonging to G. Defining the gray level co-occurrence matrix with the direction theta and the interval d as [ p (i, j, d, theta)]. Matrix [ p (i, j, d, theta)]The ith row and j column elements of (a) indicate all theta directions, one of the adjacent pixels with the interval d takes the value of i, and the other takes the adjacent pair of points of the value of i. The gray level co-occurrence matrix is normalized as follows:
p(i,j)＝p(i,j,d,θ)/R
wherein R is the normalization constant:
after the gray level co-occurrence matrix is normalized, texture characteristics can be obtained: angular second moment f1Contrast f2Correlation f3And entropy f4；
The discrete cosine transform can transform the original texture image into frequency domain to further analyze the texture characteristics of the image, the discrete cosine transform divides the whole image into N × N pixel blocks, the discrete cosine transform is carried out on the pixel blocks one by one, each block has 4 × 4 coefficients if an 8 × 8 image is averagely divided into 4 blocks, and each block uses the statistical parameter AKRepresents:
in the formula,representing the coefficients of the ith row and jth column of the nth 8 × 8 DCT block, S representing the total DCT block number of the image block, each block having 16 coefficients, Dk(k is 0,1,2,3) indicates that an 8 × 8 image is divided into 4 blocks D on average0,D1,D2,D3. Then A is0,A1,A2,A3Four feature parameters of the feature vector are constructed. In addition, EkRegion energies representing different grain directions, and:
wherein,then E0,E1,E2,E3Another 4 parameters in the feature vector are formed;
the feature vector input into the SVM classifier is then of the form f1,f2,f3,f4,A1,A2,A3,...Am,E0,E1,E2,E3。
As a further improvement of the invention, in the step 4),
the main implementation steps of SVM training are as follows:
firstly, establishing a training set N containing a positive sample and a negative sample;
secondly, training an SVM classifier by using a training set N through a nonlinear separable SVM algorithm;
thirdly, carrying out classification testing on the randomly selected non-license plate samples by using a trained SVM classifier, and collecting the samples which are wrongly classified into license plates;
fourthly, randomly selecting 50% of misclassification samples and adding the misclassification samples into the training set N;
fifthly, repeating the steps from one to four until no error sample is found;
sixthly, using the finally obtained N to train the SVM classifier;
traversing the vertical strip area obtained in the step 2) by using the detection frame in a left-right and up-down moving mode, and combining the trained SVM classifier to carry out license plate positioning, wherein the main implementation steps are as follows:
firstly, initializing the size of a detection frame to be 15 multiplied by 45 pixels;
secondly, extracting the spatial domain feature and the discrete cosine transform domain feature in the detection frame region in the vertical strip obtained in the step S2, and sending the spatial domain feature and the discrete cosine transform domain feature as input feature vectors into an SVM classifier for judgment;
if the result of the classifier is 1, marking the detection area as a license plate area, otherwise, marking the detection area as a non-license plate area;
moving the detection frame by taking 1 pixel as a step length, and recalculating the spatial domain characteristic and the discrete cosine transform domain characteristic;
and fifthly, if the traversal is finished, the detection frame is amplified by 1.25 times and traversed again, and when the height of the detection frame is larger than that of the vertical stripe or the width of the detection frame is larger than that of the vertical stripe, the traversal is finished.
As a further improvement of the present invention, the step 4) includes: combining the rectangular frames of the overlapped license plate areas according to a certain rule to form a final detection result, wherein the method for removing the overlapped rectangular frames comprises the following steps:
4.1, if the multiple rectangular frames occupy more than 50% of the respective areas in the overlapping areas, the rectangular frame of the finally detected license plate is the average value of the overlapping frames;
4.2, when the large frame surrounds the small frame, 100% of the area of the small frame is positioned in the overlapping area, and less than 50% of the area of the large frame is positioned in the overlapping area, the small frame is merged into the large frame.
Compared with the prior art, the invention has the following beneficial effects:
1. the method comprises the steps of firstly utilizing wavelet transformation to carry out multi-level wavelet decomposition on an image to obtain a scale coefficient and a wavelet coefficient, adopting an MSR (minimum shift register) method to process the scale coefficient to adjust the brightness of a low-illumination image, and adopting a threshold enhancement method to process the wavelet coefficient of each layer to enhance the image, so that a clear image with rich detail information and high contrast is obtained, and the license plate positioning is facilitated.
2. The vertical edge detection Bernsen operator is used for edge detection, the license plate is quickly roughly positioned by combining a vertical projection method, the width of the license plate region is obtained, most irrelevant background regions are filtered, and the speed of accurate positioning is favorably improved.
3. The method has the advantages that the space domain characteristics and the discrete cosine domain characteristics of the image are extracted, the characteristic quantity is rich, the license plate region can be accurately positioned by using a support vector machine method conveniently, representative negative samples are collected in a bootstrap mode in the SVM training process, a more accurate SVM classifier can be obtained, the license plate region is positioned, and then the overlapping part of the rectangular frame of the license plate region is removed to obtain the final accurate positioning effect.
Drawings
FIG. 1 is a flowchart of a license plate location method of an embodiment.
Fig. 2 is a diagram of the effect of obtaining vertical strips by rough positioning according to the embodiment.
FIG. 3 is a diagram illustrating the license plate location effect of the algorithm according to the embodiment.
Detailed Description
Example 1: the present invention will be further described in detail with reference to the accompanying drawings by taking a license plate image as an example.
S1, performing image enhancement by using wavelet transformation to obtain a clear image with rich detail information and high contrast, and facilitating accurate license plate positioning;
s1.1, performing multi-level wavelet decomposition on the image by utilizing wavelet transformation to obtain a scale coefficient and a wavelet coefficient;
let ψ (t) and φ (t) be wavelet functions and scale functions, respectively, aj[n]，dj[n]Scale coefficients and wavelet coefficients, respectively, j representing the jth wavelet decomposition,is the convolution operator. Then:
wherein:
t is the argument of the function, n is the displacement,
for any j ≧ 0, there are:
whereinIs a low-pass decomposition filter and a high-pass decomposition filter in the wavelet decomposition process.A low-pass reconstruction filter and a high-pass reconstruction filter in the corresponding wavelet inverse transformation process.
S1.2, processing the scale coefficient by adopting an MSR method to adjust the brightness of the image;
the wavelet transform has the characteristics of multi-resolution and multi-scale, when the wavelet transform is applied to an image enhancement process, the enhancement effect and the anti-noise performance are superior to those of the traditional image enhancement algorithm, but the brightness change effect of the image is poor, and aiming at the defect, the invention adopts an MSR (multiscale Retinex) method to process the low-frequency signals, namely scale coefficients of the image. The method comprises the following steps:
wherein R (x, y) is the output of the MSR process; i (x, y) is image data;is the convolution operator; w is aeAs a weight, ∑ w is satisfiede1, e denotes the order, i.e. the several weights. F (x, y) ═ k × exp (- (x)2+y2)/ce 2) K satisfies: (x, y) dxdy ═ 1; c. CeDetermining the shape of the Gaussian curved surface; e is of different shapeThe number of Gaussian curved surfaces, i.e. the number of scales, different e correspond to different ce。
Selecting three scales to process the scale coefficient of wavelet transform by MSR, cnTake the values of 15,60 and 180 respectively and weight wnTake three identical values 1/3. The MSR processes the image by utilizing the color sense consistency of human eyes, has good dynamic range compression performance on the image, and has better processing effect on the image with insufficient illumination and non-uniformity.
S1.3, processing wavelet coefficients of each layer by adopting a threshold enhancement method to enhance the detail characteristics of the image;
a length-5B-spline wavelet of 3 degrees is selected to perform a 4-level wavelet decomposition on the image. The wavelet transformation separates detail features of different resolutions in an image along with different scales, the detail features and noise of the image belong to high-frequency components, and the wavelet components of different scales are respectively transformed through different threshold algorithms according to different characteristics of wavelet coefficients of each layer, so that the detail features of different resolutions are enhanced, and the noise of the image is also inhibited.
The noise of the image belongs to high-frequency components and is mainly concentrated in a first-level wavelet coefficient, the detail characteristic of the image is enhanced by adopting a dual threshold value in the layer, and the processing method is shown as a formula (6); the noise in the wavelet coefficients of the second level and the third level is small, an adaptive threshold enhancement algorithm is adopted, and the processing method is shown as a formula (7); and (3) processing the fourth-level wavelet coefficients by adopting a single threshold algorithm, wherein the processing method is shown as an equation (8).
In the formula, T1And T2Is the threshold, T1＝3.5，G＝8，T2＝40，WinAnd WoutAre wavelet coefficients before and after the transform.
In the formula, T3And T4Is the threshold, T3＝35，G＝10，T4＝1.8，WinAnd WoutAre wavelet coefficients before and after the transform.
In the formula, T5Is the threshold, G ═ 8, T5＝1.5，WinAnd WoutAre wavelet coefficients before and after the transform.
After the processing step of S1, clear images with rich detail information and high contrast are obtained, which is beneficial to accurate license plate positioning in the follow-up process;
s2, carrying out binarization processing on the enhanced image by using a Bernsen operator, combining a vertical projection method to quickly carry out coarse positioning on the license plate to obtain the width of the license plate region, filtering out most irrelevant background regions, and facilitating the realization of subsequent SVM classification training;
after the low-illumination image is processed in the step of S1, binarization processing is performed by using a Bernsen operator, and the Bernsen operator can retain more detailed information of the image.
Since the license plate region has seven consecutive characters, there is a certain distance between the characters. There is a jump from character to background of license plate or from background to character, there are two frames in the license plate area, and there are more than two jumps between the character and the frame. The number of hops is greater than or equal to 18 for the licensed region relative to the other unlicensed regions. The license plate is generally hung at a lower position of a vehicle body, and no obvious dense edge area is arranged below the license plate, so that only the lower 2/3 part of the license plate image is vertically projected, the running speed of the algorithm is increased, and the complex background influence is reduced. The license plate is scanned from bottom to top, and the width of the license plate position is roughly positioned by combining a vertical projection method, and the method comprises the following steps:
(1) and vertically projecting the binary image processed by the Bernsen operator, scanning the license plate image after vertical projection, counting the absolute difference value of the number of black pixels in each two adjacent columns in the projected image, and simultaneously recording the coordinates of the previous column corresponding to each difference value.
(2) And arranging the absolute differences by using a quick ordering method, ordering the corresponding coordinates, taking 18 coordinate values with the maximum absolute differences, ordering from small to large, removing abrupt points by using a least square method, and obtaining the width of the license plate by using the remaining area corresponding to the coordinate values, namely the area of the license plate in the vertical direction, namely the area of a vertical strip.
The width of the license plate is obtained, and meanwhile, most irrelevant complex background areas are filtered, so that the realization of subsequent SVM classification training is facilitated. Fig. 2 is a diagram of the effect of obtaining vertical strips by coarse positioning.
S3, extracting the spatial domain characteristics and the discrete cosine transform domain characteristics of the image, and performing classification training by using an SVM algorithm so as to position a license plate region;
s3.1, extracting the spatial domain characteristics and the discrete cosine transform domain characteristics of the image to be used as the characteristic vector of SVM training;
the space domain characteristics and the characteristics of the discrete cosine transform domain of the license plate region obtained in the step S2 are extracted, the characteristic quantity is rich, and the support vector machine is convenient to accurately position the license plate. Extracting image surface texture characteristic parameters by adopting a gray level co-occurrence matrix for spatial domain characteristics, grading the gray level value of the image f (i, j), wherein the highest gray level is NgThe width of the image in the horizontal direction is Lx＝{1,2,...,NxLength of image in vertical direction is Ly＝{1,2,...,NyAnd recording a gray level set of the image as G ═ 1,2g}. The image f (i, j) to be analyzed can be considered as Lx×LyOne transformation to G, i.e. each point L in the image f (i, j)x×LyCorresponding to a gray value belonging to G. Defining the gray level co-occurrence matrix with the direction theta and the interval d as [ p (i, j, d, theta)]. Matrix [ p (i, j, d, theta)]The ith row and j column elements of (a) indicate all theta directions, one of the adjacent pixels with the interval d takes the value of i, and the other takes the adjacent pair of points of the value of i. The gray level co-occurrence matrix is normalized as follows:
p(i,j)＝p(i,j,d,θ)/R (9)
wherein R is the normalization constant:
after the gray level co-occurrence matrix is normalized, texture characteristics can be obtained: angular second moment, contrast, correlation and entropy. The angular second moment describes the characteristic of uniform distribution of image gray, the value of the coarse texture is larger, and the fine texture is smaller, as shown in formula (11); the contrast describes the definition of the image, and the deeper the texture groove, the higher the contrast, the clearer the image, as shown in formula (12); the correlation is used to measure the similarity of the elements of the gray level co-occurrence matrix in the row direction or the column direction, as shown in formula (13); the entropy is used to measure the information content of the image, and if the image has no texture, the entropy value is almost 0, and the formula of the entropy is shown in formula (16).
In the formula, mux，σxAre each { px(i)；i＝1,2,...,NgMean and mean square error of }; mu.sy，σyAre each { py(i)；i＝1,2,...,NgMean and mean square error of. Wherein:
the above features will yield different values for different pitches d and directions θ, typically given that d is 1 and θ is 0 °, 45 °, 90 °, 135 °, so for each feature fiThere are 4 values in different directions (i ═ 1,2,3, 4).
The discrete cosine transform is to divide the whole image into N × N pixel blocks, and then to make discrete cosine transform one by one, and to assume that an 8 × 8 image is divided into 4 blocks on average, each block has 4 × 4 coefficients, and each block uses the statistical parameter A to analyze the blockKRepresents:
in the formula,representing the coefficients of the ith row and jth column of the nth 8 × 8 DCT block, S representing the total DCT block number of the image block, each block having 16 coefficients, Dk(k is 0,1,2,3) indicates that an 8 × 8 image is divided into 4 blocks D on average0,D1,D2,D3. Then A is0,A1,A2,A3Four feature parameters of the feature vector are constructed. In addition, EkRegion energies representing different grain directions, and:
wherein,then E0,E1,E2,E3The other 4 parameters in the feature vector are constructed.
The feature vector input into the SVM classifier is then of the form f1,f2,f3,f4,A1,A2,A3,...Am,E0,E1,E2,E3。
S3.2SVM algorithm is used for classification training to locate the license plate area.
The SVM constructs an optimal hyperplane based on the SRM criterion, so that the interval between each type of data is maximum, and meanwhile, the classification error is kept as small as possible. The problem of accurate positioning of a license plate is a nonlinear separable mode classification problem. Given a training set (x)1,y1)，(x2,y2)，…,(xi,yi) Wherein x isiAs an input vector, yiIs a sample label, yi± 1. In the present inventionyi± 1. For the linear separable problem, the training samples can be classified into 2 classes according to the hyperplane shown in equation (20):
yi(xi·w+b)≥1,i＝1,2,...,m (20)
where w is the vector in the hyperplane and b is the intercept.
The edge of the hyperplane is the sum of the shortest distances from the hyperplane to the positive and negative examples, as follows:
where M is the edge of the hyperplane, x+Is a support vector of positive samples, x-Is the support vector for the negative examples and g is the normalized vector for the hyperplane. The problem of solving the maximum value of M in the SVM is converted into solving | | w | | luminance2Is the minimum value of (a), formula (20) is its constraint.
For the linear inseparable license plate location problem, the slack variable ξ can be usediAnd a penalty factor C is introduced into each training dataset. The formula (21) is converted into the formula (22).
The constraint conditions are as follows:
finding the optimal hyperplane can be classified as a quadratic programming problem:
and mapping the training set to a high-dimensional feature space, and replacing dot product operation with a kernel function. In the high-dimensional feature space, the data is linearly separable. The quadratic programming problem (24) is converted to solve the maximum lagrangian:
constraint conditions are as follows:
in the formula, αiAnd αjAre Lagrangian dual variables, x and y are as defined above, K (x)i,xj) For the kernel function, a Gaussian kernel function K (x) is used in this casei,xj)＝exp(-1/(2σ2)||xi-xj||2) And σ is the bandwidth of the kernel function.
At the time of obtaining αiAfter the optimal solution of (2), the classification of the test sample x can be judged by the following expression (27).
In the process of positioning the license plate by using the SVM algorithm, the image containing the license plate area can be used as a positive sample, and the difficulty is how to collect representative negative samples, because in fact, too many images not containing the license plate can be used as negative samples. In the method, some negative samples are obtained in training instead of being selected before training, so that a more accurate SVM classifier can be obtained, the classification is more accurate, and the license plate positioning effect is more accurate. The main implementation steps of SVM training are as follows:
(1) establishing a training set N containing a positive sample and a negative sample;
(2) training an SVM classifier by using a training set N through a nonlinear separable SVM algorithm;
(3) carrying out classification test on the randomly selected non-license plate samples by using a trained SVM classifier, and collecting the samples which are wrongly classified into license plates;
(4) randomly selecting 50% misclassification samples and adding the misclassification samples into a training set N;
(5) repeating the steps (2) to (4) until no misclassified sample is found;
(6) and training the SVM classifier by using the finally obtained N.
Traversing the vertical strip area obtained in the step S2 by using a detection frame with a certain size in a left-right and up-down moving mode, and performing license plate positioning by combining a trained SVM classifier, wherein the method mainly comprises the following steps:
(1) initializing the detection frame size to be 15 × 45 pixels;
(2) extracting the spatial domain feature and the discrete cosine transform domain feature in the detection frame region in the vertical strip obtained in the step S2, taking the spatial domain feature and the discrete cosine transform domain feature as input feature vectors, and sending the input feature vectors into an SVM classifier for judgment;
(3) if the result of the classifier is 1, marking the detection area as a license plate area, otherwise, marking the detection area as a non-license plate area;
(4) moving the detection frame by taking 1 pixel as a step length, and recalculating the spatial domain characteristic and the discrete cosine transform domain characteristic;
(5) and if the traversal is finished, amplifying the detection frame by 1.25 times, and traversing again, wherein the traversal is finished when the height of the detection frame is larger than the height of the vertical stripe, or the width of the detection frame is larger than the width of the vertical stripe.
And S4, detecting the overlapped rectangular frame of the license plate area, and accurately positioning the license plate area.
The license plate region located in step S3 may have multiple overlapping windows. In order to accurately locate the license plate region, the rectangular frames of the overlapped license plate regions must be combined according to a certain rule to form a final detection result. The method for removing the overlapped rectangular frames comprises the following steps:
(1) if the multiple rectangular frames occupy more than 50% of the respective areas in the overlapping areas, the rectangular frame of the license plate finally detected is the average value of the overlapping frames;
(2) when the large frame surrounds the small frame, 100% of the area of the small frame is located in the overlapping area, and less than 50% of the area of the large frame is located in the overlapping area, the small frame is fused into the large frame.
Fig. 3 is a diagram illustrating the effect of the algorithm of the present embodiment on license plate location. The image result shows that the algorithm of the embodiment has good effect of positioning the license plate under low illumination. For the license plate image with low illumination, the algorithm based on the SVM algorithm and the algorithm of the embodiment is respectively adopted for license plate positioning, and the comparison of the detection results is shown in the following table:
TABLE 1 comparison of test results
From the above table, it can be seen that the performance of the algorithm of the embodiment is better than that of the algorithm based on the SVM only, because the candidate area is greatly reduced after the vertical strips are obtained by coarse positioning, the number of the fake license plates is reduced, and the accuracy is improved. The algorithm of the embodiment can meet the real-time requirement of engineering in time.
The method provided by the invention can be actually embedded into an FPGA (field programmable gate array) to realize, and is applied to a license plate recognition and video monitoring system. The above embodiments only serve to explain the technical solution of the present invention, and the protection scope of the present invention is not limited to the implementation system and the specific implementation steps described in the above embodiments. Therefore, the technical solutions that the specific formulas and algorithms in the above embodiments are simply replaced, but the substantial contents are still consistent with the method of the present invention, and all the technical solutions are within the protection scope of the present invention.

Claims (7)

1. A license plate positioning method based on wavelet transformation and SVM low-illumination complex background is characterized by comprising the following steps:
step 1) image enhancement is carried out on the low-illumination image by utilizing wavelet transformation to obtain a high-contrast image, and the method comprises the following steps:
1.1, performing multi-level wavelet decomposition on the image by utilizing wavelet transformation to obtain a scale coefficient phi (t) and a wavelet coefficient psi (t);
1.2, processing the scale coefficient by adopting an MSR method to adjust the brightness of the image;
1.3, processing wavelet coefficients of each layer by adopting a threshold enhancement method to enhance the detail characteristics of the image;
step 2) carrying out binarization processing on the enhanced image, and combining the binarized image with a vertical projection method to quickly carry out coarse positioning on the license plate so as to obtain the width of the license plate region;
step 3) extracting the airspace characteristics and discrete cosine domain characteristics of the license plate region obtained in the step 2), and performing classification training by using an SVM algorithm so as to locate the license plate region, wherein the method comprises the following steps:
3.1, extracting the spatial domain characteristic and the discrete cosine characteristic of the image to be used as a characteristic vector of SVM training;
3.2, carrying out classification training by utilizing an SVM algorithm, and positioning a license plate area;
and 4) detecting the overlapped rectangular frames of the license plate area, and accurately positioning the license plate area.
2. The method for locating the license plate with the complex background and low illumination based on the wavelet transform and the SVM of claim 1, wherein in the step 1), the method for adjusting the brightness of the image by processing the scale coefficient by the MSR method is as follows:
R ( x , y ) = Σ e = 1 E w e { log I ( x , y ) - l o g [ F ( x , y ) ⊗ I ( x , y ) ] }
in the above formula, R (x, y) is the output of the MSR method; i (x, y) is image data;is the convolution operator; w is aeAs a weight, ∑ w is satisfiede1, e denotes the order, i.e. the number of weights, F (x, y) K x exp (- (x)2+y2)/ce 2) K satisfies: (x, y) dxdy ═ 1; c. CeDetermining the shape of the Gaussian curved surface; e is the number of Gaussian curved surfaces with different shapes, namely the number of scales, and different E correspond to different ce。
3. The method for locating the license plate with the complex background and low illumination based on the wavelet transform and the SVM of claim 1, wherein in the step 1), the step of processing wavelet coefficients of each layer by adopting a threshold enhancement method to enhance the detail characteristics of the image comprises the following steps:
selecting 3-time B-spline wavelet with the length of 5 to carry out four-level wavelet decomposition on the image, and respectively transforming wavelet components with different scales through different threshold algorithms according to different characteristics of wavelet coefficients of each layer;
within the first-level wavelet coefficient, image detail features are enhanced by adopting a dual threshold:
W o u t = W i n + ( T 2 * ( G - 1 ) ) - ( T 1 * G ) , W i n > T 2 G * ( W i n - T 1 ) , T 2 &GreaterEqual; W i n > T 1 0 , - T 1 &le; W i n &le; T 1 G * ( W i n + T 1 ) , - T 1 > W i n &GreaterEqual; - T 2 W i n - ( T 2 * ( G - 1 ) ) + ( T 1 * G ) , W i n < - T 2
in the above formula, T1And T2Is the threshold, T1＜T2，G＝8，WinAnd WoutIs the wavelet coefficient before and after transformation;
in the second and third wavelet coefficients, adopting an adaptive threshold enhancement algorithm:
W o u t = W i n , W i n > T 3 T 3 - G * T 4 T 3 - T 4 * ( W i n - T 4 ) + T 3 , T 3 &GreaterEqual; W i n > T 4 G * W i n , - T 4 &le; W i n &le; T 4 T 3 - G * T 4 T 3 - T 4 * ( W i n + T 4 ) - T 3 , - T 4 > W i n &GreaterEqual; - T 3 W i n , W i n < - T 3
in the formula, T3And T4Is the threshold, T4＜T3，G＝10，WinAnd WoutIs the wavelet coefficient before and after transformation;
and (3) in a fourth-level wavelet coefficient, processing by adopting a single threshold algorithm:
W o u t = W i n + T 5 ( G - 1 ) , W i n > T 5 W i n - T 5 ( G - 1 ) , W i n < - T 5 G * W i n , - T 5 < W i n < T
in the formula, T5Is the threshold, G ═ 8, WinAnd WoutAre wavelet coefficients before and after the transform.
4. The wavelet transform and SVM based low-illumination complex background license plate positioning method of claim 1, wherein in the step 2), the lower 2/3 part of the license plate image is vertically projected, the license plate is scanned in a bottom-up manner, and the width of the license plate position is roughly positioned by combining the vertical projection method comprises the following steps:
(1) vertically projecting the processed binary image, scanning the vertically projected license plate image, counting the absolute difference of the number of black pixel points of two adjacent columns in the projected image, and simultaneously recording the coordinates of the previous column corresponding to each absolute difference;
(2) and arranging the absolute differences by using a quick ordering method, ordering the corresponding coordinates, taking 18 coordinate values with the maximum absolute differences, ordering from small to large, removing abrupt points by using a least square method, and obtaining the width of the license plate by using the remaining area corresponding to the coordinate values, namely the area of the license plate in the vertical direction, namely the area of the vertical strip.
5. The method for locating license plate with low-illumination complex background according to claim 1, wherein in step 3), a gray level co-occurrence matrix is used for spatial domain features to extract image surface texture feature parameters, the gray level of the image f (i, j) is graded, and the highest gray level is NgThe width of the image in the horizontal direction is Lx＝{1,2,...,NxLength of image in vertical direction is Ly＝{1,2,...,NyAnd recording a gray level set of the image as G ═ 1,2gThe image f (i, j) to be analyzed can be considered as Lx×LyOne transformation to G, i.e. each point L in the image f (i, j)x×LyCorresponding to a gray value belonging to G, defining a gray co-occurrence matrix with a direction theta and an interval d as [ p (i, j, d, theta)]Matrix [ p (i, j, d, theta)]The ith row and j column elements of the gray level co-occurrence matrix represent all theta directions, one of the pixels with adjacent interval d takes the value of i, the other takes the number of adjacent pairs of points of the value of i, and the normalization processing of the gray level co-occurrence matrix is as follows:
p(i,j)＝p(i,j,d,θ)/R
wherein R is the normalization constant:
after the gray level co-occurrence matrix is normalized, texture characteristics can be obtained: angular second moment f1Contrast f2Correlation f3And entropy f4；
The discrete cosine transform can transform the original texture image into the frequency domain to further analyze the texture characteristics of the image, and the discrete cosine transform divides the whole image into N × NA pixel block, which is subjected to discrete cosine transform one by one, wherein if an 8 × 8 image is divided into 4 blocks on average, each block has 4 × 4 coefficients, and each block uses the statistical parameter AKRepresents:
A k = 1 16 S &Sigma; n = 1 S &Sigma; D k C i j n , k = 0 , 1 , 2 , 3
in the formula,representing the coefficients of the ith row and jth column of the nth 8 × 8 DCT block, S representing the total DCT block number of the image block, each block having 16 coefficients, Dk(k is 0,1,2,3) indicates that an 8 × 8 image is divided into 4 blocks D on average0,D1,D2,D3Then A is0,A1,A2,A3Four feature parameters, E, forming a feature vectorkRegion energies representing different grain directions, and:
E 0 = &Sigma; n = 1 S ( C 00 n - C &OverBar; 0 ) 2
E k = &Sigma; n = 1 &Sigma; D k ( C i j n - C &OverBar; k ) 2 , k = 0 , 1 , 2 , 3
wherein,then E0,E1,E2,E3Another 4 parameters in the feature vector are formed;
the feature vector input into the SVM classifier is then of the form f1,f2,f3,f4,A1,A2,A3,...Am,E0,E1,E2,E3。
6. The method for locating the license plate with the complex background and low illumination based on the wavelet transform and SVM of claim 1, wherein in the step 4),
the main implementation steps of SVM training are as follows:
firstly, establishing a training set N containing a positive sample and a negative sample;
secondly, training an SVM classifier by using a training set N through a nonlinear separable SVM algorithm;
thirdly, carrying out classification testing on the randomly selected non-license plate samples by using a trained SVM classifier, and collecting the samples which are wrongly classified into license plates;
fourthly, randomly selecting 50% of misclassification samples and adding the misclassification samples into the training set N;
fifthly, repeating the steps from one to four until no error sample is found;
sixthly, using the finally obtained N to train the SVM classifier;
traversing the vertical strip area obtained in the step 2) by using the detection frame in a left-right and up-down moving mode, and combining the trained SVM classifier to carry out license plate positioning, wherein the main implementation steps are as follows:
firstly, initializing the size of a detection frame to be 15 multiplied by 45 pixels;
secondly, extracting the spatial domain feature and the discrete cosine transform domain feature in the detection frame region in the vertical strip obtained in the step S2, and sending the spatial domain feature and the discrete cosine transform domain feature as input feature vectors into an SVM classifier for judgment;
if the result of the classifier is 1, marking the detection area as a license plate area, otherwise, marking the detection area as a non-license plate area;
moving the detection frame by taking 1 pixel as a step length, and recalculating the spatial domain characteristic and the discrete cosine transform domain characteristic;
and fifthly, if the traversal is finished, the detection frame is amplified by 1.25 times and traversed again, and when the height of the detection frame is larger than that of the vertical stripe or the width of the detection frame is larger than that of the vertical stripe, the traversal is finished.
7. The wavelet transform and SVM based low-illumination complex background license plate location method according to claim 1, wherein the step 4) comprises the steps of: combining the rectangular frames of the overlapped license plate areas according to a certain rule to form a final detection result, wherein the method for removing the overlapped rectangular frames comprises the following steps:
4.1, if the multiple rectangular frames occupy more than 50% of the respective areas in the overlapping areas, the rectangular frame of the finally detected license plate is the average value of the overlapping frames;
4.2, when the large frame surrounds the small frame, 100% of the area of the small frame is positioned in the overlapping area, and less than 50% of the area of the large frame is positioned in the overlapping area, the small frame is merged into the large frame.
CN201610976425.5A 2016-11-07 2016-11-07 Low-illumination complicated background license plate positioning method based on wavelet transform and SVM Pending CN106407959A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201610976425.5A CN106407959A (en) 2016-11-07 2016-11-07 Low-illumination complicated background license plate positioning method based on wavelet transform and SVM

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201610976425.5A CN106407959A (en) 2016-11-07 2016-11-07 Low-illumination complicated background license plate positioning method based on wavelet transform and SVM

Publications (1)

Publication Number Publication Date
CN106407959A true CN106407959A (en) 2017-02-15

Family

ID=58015454

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201610976425.5A Pending CN106407959A (en) 2016-11-07 2016-11-07 Low-illumination complicated background license plate positioning method based on wavelet transform and SVM

Country Status (1)

Country Link
CN (1) CN106407959A (en)

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108776792A (en) * 2018-06-07 2018-11-09 北京智芯原动科技有限公司 A kind of multiple dimensioned positioning fusion method and device of car plate
CN110097132A (en) * 2019-05-07 2019-08-06 电子科技大学 A method of identification digital photos and shooting camera
CN110163815A (en) * 2019-04-22 2019-08-23 桂林电子科技大学 Low-light (level) restoring method based on multistage variation self-encoding encoder
CN112216640A (en) * 2020-10-19 2021-01-12 惠州高视科技有限公司 Semiconductor chip positioning method and device
CN113903180A (en) * 2021-11-17 2022-01-07 四川九通智路科技有限公司 Method and system for detecting vehicle overspeed on expressway
CN113984192A (en) * 2021-10-27 2022-01-28 广东电网有限责任公司佛山供电局 Transformer working state monitoring method and system based on sound signals

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101246551A (en) * 2008-03-07 2008-08-20 北京航空航天大学 Fast license plate locating method
CN102831429A (en) * 2012-08-13 2012-12-19 深圳市捷顺科技实业股份有限公司 License plate locating method
CN103824078A (en) * 2014-03-18 2014-05-28 厦门翼歌软件科技有限公司 Complex scene multi-license plate positioning method

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101246551A (en) * 2008-03-07 2008-08-20 北京航空航天大学 Fast license plate locating method
CN102831429A (en) * 2012-08-13 2012-12-19 深圳市捷顺科技实业股份有限公司 License plate locating method
CN103824078A (en) * 2014-03-18 2014-05-28 厦门翼歌软件科技有限公司 Complex scene multi-license plate positioning method

Non-Patent Citations (6)

* Cited by examiner, † Cited by third party
Title
于明 等: "自适应复杂天气的车牌定位方法", 《计算机工程与科学》 *
刘尚旺 等: "二次定位车牌分割及识别方法", 《河南师范大学学报（自然科学版）》 *
周旋 等: "基于小波变换的图像增强新算法", 《计算机应用》 *
贺光: "基于粗糙集和模糊SVM的车牌识别技术研究", 《中国优秀硕士学位论文全文数据库-信息科技辑》 *
郭延祥: "针对高分辨率背景复杂图像的车牌定位算法研究", 《中国优秀硕士学位论文全文数据库-信息科技辑》 *
隆晓玲 等: "基于SVM的车牌区域定位系统研究", 《信息技术》 *

Cited By (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108776792A (en) * 2018-06-07 2018-11-09 北京智芯原动科技有限公司 A kind of multiple dimensioned positioning fusion method and device of car plate
CN108776792B (en) * 2018-06-07 2021-09-17 北京智芯原动科技有限公司 Multi-scale positioning fusion method and device for license plate
CN110163815A (en) * 2019-04-22 2019-08-23 桂林电子科技大学 Low-light (level) restoring method based on multistage variation self-encoding encoder
CN110163815B (en) * 2019-04-22 2022-06-24 桂林电子科技大学 Low-illumination reduction method based on multi-stage variational self-encoder
CN110097132A (en) * 2019-05-07 2019-08-06 电子科技大学 A method of identification digital photos and shooting camera
CN112216640A (en) * 2020-10-19 2021-01-12 惠州高视科技有限公司 Semiconductor chip positioning method and device
CN112216640B (en) * 2020-10-19 2021-08-06 高视科技(苏州)有限公司 Semiconductor chip positioning method and device
CN113984192A (en) * 2021-10-27 2022-01-28 广东电网有限责任公司佛山供电局 Transformer working state monitoring method and system based on sound signals
CN113984192B (en) * 2021-10-27 2023-08-01 广东电网有限责任公司佛山供电局 Transformer working state monitoring method and system based on sound signals
CN113903180A (en) * 2021-11-17 2022-01-07 四川九通智路科技有限公司 Method and system for detecting vehicle overspeed on expressway

Similar Documents

Publication Publication Date Title
CN115861135B (en) Image enhancement and recognition method applied to panoramic detection of box body
CN106407959A (en) Low-illumination complicated background license plate positioning method based on wavelet transform and SVM
JP6710135B2 (en) Cell image automatic analysis method and system
CN110443128B (en) Finger vein identification method based on SURF feature point accurate matching
CN103048329B (en) A kind of road surface crack detection method based on active contour model
CN107103317A (en) Fuzzy license plate image recognition algorithm based on image co-registration and blind deconvolution
CN109635733B (en) Parking lot and vehicle target detection method based on visual saliency and queue correction
CN108765465A (en) A kind of unsupervised SAR image change detection
US20240221201A1 (en) Linewidth measurement mthod and apparatus, computing and processing device, computer program and computer readable medium
CN105279772A (en) Trackability distinguishing method of infrared sequence image
CN108932518A (en) A kind of feature extraction of shoes watermark image and search method of view-based access control model bag of words
CN106557740A (en) The recognition methods of oil depot target in a kind of remote sensing images
CN105894037A (en) Whole supervision and classification method of remote sensing images extracted based on SIFT training samples
CN109507193A (en) A kind of fabric defects detection method based on local contrast enhancing and binary pattern
CN104637060B (en) A kind of image partition method based on neighborhood principal component analysis-Laplce
CN107784284B (en) Face recognition method and system
CN103942526A (en) Linear feature extraction method for discrete data point set
CN112990368B (en) Polygonal structure guided hyperspectral image single sample identification method and system
CN115147613A (en) Infrared small target detection method based on multidirectional fusion
CN111275687B (en) Fine-grained image stitching detection method based on connected region marks
CN106295478A (en) A kind of image characteristic extracting method and device
CN109754001B (en) Image classification method, computer storage medium, and image classification device
Bojarczak Visual algorithms for automatic detection of squat flaws in railway rails
Tung et al. Efficient uneven-lighting image binarization by support vector machines
CN112818779B (en) Human behavior recognition method based on feature optimization and multiple feature fusion

Legal Events

Date Code Title Description
C06 Publication
PB01 Publication
C10 Entry into substantive examination
SE01 Entry into force of request for substantive examination
RJ01 Rejection of invention patent application after publication

Application publication date: 20170215

RJ01 Rejection of invention patent application after publication