CN110489587A - The tire trace image characteristic extracting method of three value mode of Local gradient direction - Google Patents

Abstract

A tire trace image feature extraction method in local gradient direction ternary mode is composed of image preprocessing, feature extraction, determination of feature vector, and image retrieval steps. The invention proposes a local gradient direction ternary pattern feature suitable for tire track images, uses a more stable gradient direction value instead of gray value to perform local texture encoding, and performs threshold quantization on the gradient direction angle of the central pixel to generate high-quality texture. The edge information improves the retrieval accuracy; the Manhattan distance is used to calculate the similarity of the features described by the feature vector, and the retrieval results are obtained, and the retrieval accuracy is significantly better than other texture features. It has the advantages of clear texture edge information of tire track images, high retrieval accuracy, high average precision, and is suitable for large sample data, and can be used for feature extraction of tire track images.

Description

Tire Track Image Feature Extraction Method Based on Ternary Mode of Local Gradient Direction

technical field

The invention belongs to the technical field of image processing, and specifically relates to an image feature extraction method.

Background technique

Usually in a traffic accident, the relationship between the tire indentation marks left on the scene and the physical evidence will be used to explain the process of the accident and judge the responsibility of both parties. The tire traces on the ground are often one of the most useful traces of physical evidence, so tire trace retrieval is often used in public security. Clue acquisition in case solving or traffic accident handling.

The research related to tire mark image retrieval in my country started late, and the research on tire marks is even less fruitful, and there is no standard tire mark image test database in the field of tire mark image retrieval and classification research. At present, the classification and retrieval of tire track images is mainly carried out by the method of combining traditional features, and there is no systematic analysis of the characteristics of tire track images. For example, by analyzing the principle of SIFT transform and Gabor wavelet, a pattern recognition method of tire tracks based on SIFT-Gabor transform is proposed. Combining unsupervised learning and hierarchical feature extraction, a tire track image retrieval method based on sparse representation and probabilistic latent semantic analysis is proposed. A combined feature extraction method based on Non-Subsampled Contourlet Transform (NSCT) and Gray Level Co-occurrence matrix (GLCM), and uses a multi-level support vector machine (Support Vector Machine, SVM) to train the classifier. In this method, the tire track image formed by the tire and the ground includes the basic features of the main line groove and the edge grain line through analysis. At the same time, the edge grain lines are very complex. The edge grain lines of different types of tires have different widths and directions, and the grain lines are interconnected and intertwined, and the trace color is relatively single.

Among various texture image descriptors, Local Binary Patterns (LBP) is a popular and powerful image descriptor. It has low computational complexity, no training and learning, and easy engineering implementation, so it has received extensive attention in the fields of computer vision and image processing. Based on the original LBP method, scholars have proposed many improved LBP methods. For example: Local Directional Patterns (LDP). Enhanced Local Directional Patterns (Enhanced Local Directional Patterns, ELDP. Local Directional Number (LDN), etc.

In the field of image processing technology, a technical problem that needs to be urgently solved at present is to provide a method for extracting the image features of tire tracks.

SUMMARY OF THE INVENTION

The technical problem to be solved by the present invention is to overcome the above-mentioned shortcomings of the prior art, and to provide a local gradient direction ternary model with clear tire trace image texture edge information, high retrieval accuracy, high average precision, and suitable for large sample data. A method for feature extraction of tire track images.

The technical solution adopted to solve the above-mentioned technology is composed of the following steps:

(1) Image preprocessing

From the tire track image database, 50 to 80 tire track sample images of 30 categories are selected for size normalization and grayscale processing.

(2) Feature extraction

1) Use the Sobel edge detection method to determine the image difference G _x along the x direction and the image difference G _y in the y direction, and use the formula (1) to determine the gradient direction angle α(x, y) of each pixel in the image

α(x,y)=arctan(G _y /G _x ) (1)

2) In the 3×3 neighborhood sliding window, use the local gradient direction ternary mode LGDTP method to determine each gradient direction angle value, and the local gradient direction ternary mode LGDTP method increases the custom threshold t, g _i is greater than the interval [g _c -t,g _c +t]} is 1, belongs to this interval is 0, is smaller than this interval is -1, and obtains a three-valued encoded value.

where P is the number of pixels in the neighborhood, R is the radius of the neighborhood, 0 < t < 2π, g _c is the gradient direction angle of the central pixel, and _gi is the gradient direction angle of its neighborhood pixels.

3) Decompose the LGDTP eigenvalue into positive and negative coding eigenvalues, modify the coding value to 0 if the coding value is not 1, obtain the positive coding eigenvalue LGDTP _P , modify the coding value -1 to 1, and modify the remaining coding values to 0 to obtain the negative coding Eigenvalue LGDTP _M .

4) The positive coded feature value LGDTP _P replaces the pixel values in the image to construct a positive coded image, and the negative coded feature value LGDTP _M replaces the pixel values in the image to construct a negative coded image.

(3) Determine the feature vector

1) Divide the positive coded image composed of the positive coding eigenvalue LGDTP _P and the negative coded image composed of the negative coding eigenvalue LGDTP _M into 3×3 sub-blocks uniformly, from top to bottom in order, and labelled 1 from left to right ~9 sub-regions, the pixels of each sub-region are m×n, and m and n are 128 or 256.

2) Perform histogram statistics on the gradient direction angles of the pixels in each sub-block of the positive-coded image, and cascade all the sub-block histograms; perform histogram statistics on the gradient direction angles of the pixels in each sub-block in the negative-coded image, level Concatenate all sub-block histograms, and concatenate the sub-block histograms of the positive and negative coded images.

3) The concatenated histogram is used as the feature of the entire tire track image, the length of the features of the positive encoding image LGDTP _M and the negative encoding image LGDTP _P are 256 respectively, and the feature length of the final encoded image LGDTP is 256 × 2 × 5, Denoted as F _LGDTP .

4) Normalize the F _LGDTP feature according to formula (4) to obtain the normalized eigenvector value:

where F ^c (t) is the t-th component of feature c, is the normalized eigenvector value.

(4) Image retrieval

1) Determine the eigenvector value of each tire track image Take each image as a query image and each image in the tire track image database to measure the similarity using the Manhattan distance metric method according to (5):

where d is the length of the distance between the features of two tire track images, X _i , X _j represent the feature vector of each tire track image;

2) Using the average precision rate P as the retrieval performance evaluation index, it is determined by the following formula (6):

Where S is the number of correct images included in the query result, K is the total number of images in the query result, and K is less than the number of sheets of each category in the tire track image database.

In the step 1) of the step (3) of determining the feature vector of the present invention, the m and n are positive integers that are divisible by 3 and are equal.

In step 2) of the feature extraction step (2) of the present invention, t in formula (3) is preferably π/6.

In the step 2) of the step (3) of determining the feature vector of the present invention, according to the particularity of the spatial distribution of the texture information of the tire track image, the sub-block histograms are labeled 2, 4, 5, 6, and 8 respectively. Histogram of the area.

The invention extracts the image features according to the tire track image database image and the query image, and uses the query image to query the tire track image database of related types of tire track images to determine the tire type. Through the analysis, it is concluded that the tire track image formed by the tire and the ground includes the basic features of the main line groove and the edge texture line. The invention proposes a local gradient direction ternary pattern feature suitable for tire track images, uses a more stable gradient direction value instead of gray value to perform local texture coding, and quantizes the gradient direction angle of the central pixel by thresholding, producing high-quality images. The texture edge information improves the retrieval accuracy; the Manhattan distance is used to calculate the similarity of the features described by the feature vector, and the retrieval results are obtained, and the retrieval accuracy is significantly better than other texture features. The invention has the advantages of clear texture edge information of tire track images, high retrieval accuracy, high average precision, and is suitable for large sample data, and can be used for feature extraction of tire track images.

Description of drawings

FIG. 1 is a flow chart of Embodiment 1 of the present invention.

FIG. 2 is a comparison experiment result diagram of the method of Embodiment 1 and 6 existing texture feature extraction methods.

FIG. 3 is a comparative experiment curve between the method of Embodiment 1 and 6 existing texture feature extraction methods.

Detailed ways

The present invention will be further described in detail below with reference to the accompanying drawings and examples, but the present invention is not limited to the following embodiments.

Example 1

The images in this example come from the tire track image database built by the applicant, including 30 categories of 80 images in each category, a total of 2,400 images. Experiments have been carried out. The steps of the tire track image feature extraction method in the local gradient direction ternary mode are as follows (see Fig. 1):

(1) Image preprocessing

From the tire track image database, select 2400 tire track sample images in 30 categories, 80 for each category, and normalize the size to 384 × 384 and grayscale.

(2) Feature extraction

1) Use the Sobel edge detection method to determine the image difference G _x along the x direction and the image difference G _y in the y direction, and use the formula (1) to determine the gradient direction angle α(x, y) of each pixel in the image

α(x,y)=arctan(G _y /G _x ) (7)

2) In the 3×3 neighborhood sliding window, use the local gradient direction ternary mode LGDTP method to determine each gradient direction angle value, and the local gradient direction ternary mode LGDTP method adds a custom threshold t, g _i is greater than the interval [g _c -t,g _c +t]} is 1, belongs to this interval is 0, is smaller than this interval is -1, and obtains a three-valued encoded value:

where P is the number of pixels in the neighborhood, R is the radius of the neighborhood, t in this embodiment is 2π, g _c is the gradient direction angle of the central pixel, and _gi is the gradient direction angle of its neighborhood pixels.

3) Decompose the LGDTP eigenvalue into positive and negative coding eigenvalues, modify the coding value to 0 if the coding value is not 1, obtain the positive coding eigenvalue LGDTP _P , modify the coding value -1 to 1, and modify the remaining coding values to 0 to obtain the negative coding Eigenvalue LGDTP _M .

4) The positive coded feature value LGDTP _P replaces the pixel values in the image to construct a positive coded image, and the negative coded feature value LGDTP _M replaces the pixel values in the image to construct a negative coded image.

(3) Determine the feature vector

1) Divide the positive coded image composed of the positive coding eigenvalue LGDTP _P and the negative coded image composed of the negative coding eigenvalue LGDTP _M into 3×3 sub-blocks uniformly, from top to bottom in order, and labelled 1 from left to right ~9 sub-regions, the pixels of each sub-region are m×n, and m and n in this embodiment are 128.

2) Perform histogram statistics on the gradient direction angles of the pixels in each sub-block of the positive-coded image, and cascade all the sub-block histograms; perform histogram statistics on the gradient direction angles of the pixels in each sub-block in the negative-coded image, level Concatenate all sub-block histograms, and concatenate the sub-block histograms of the positive and negative coded images. According to the particularity of the spatial distribution of the texture information of the tire track image, the above-mentioned sub-block histograms are histograms of sub-regions numbered 2, 4, 5, 6, and 8 respectively.

3) The concatenated histogram is used as the feature of the entire tire track image, the length of the features of the positive encoding image LGDTP _M and the negative encoding image LGDTP _P are 256 respectively, and the feature length of the final encoded image LGDTP is 256 × 2 × 5, Denoted as F _LGDTP .

4) Normalize the F _LGDTP feature according to formula (10) to obtain the normalized eigenvector value:

where F ^c (t) is the t-th component of feature c, is the normalized eigenvector value.

(4) Image retrieval

1) Determine the eigenvector value of each tire track image Take each image as a query image and each image in the tire track image database to measure the similarity using the Manhattan distance metric method according to (11):

where d is the length of the distance between the features of two tire track images, and X _i and X _j represent the feature vector of each tire track image.

2) Using the average precision rate P as the retrieval performance evaluation index, it is determined by the following formula (12):

Where S is the number of correct images contained in the query result, and K is the total number of images in the query result.

In this embodiment, the method of this embodiment and the local binary mode LBP method, the local direction mode LDP method, the enhanced local direction mode ELDP method, the local direction number mode LDN method, the optimized local direction mode OLDP method, and the local three-dimensional mode are adopted. The value mode LTP method uses the Manhattan distance to carry out comparative retrieval experiments in the tire track image database, the local binary mode LBP method, the local direction mode LDP method, the enhanced local direction mode ELDP method, the local direction number mode LDN method, and the optimized local direction mode. The feature dimension of the _OLDP method is 256×5, and the LTP method of the local three-value mode is to decompose the LTP eigenvalues into positive and negative coded eigenvalues. _M replaces the pixel values in the image to construct a negative coded image, the positive coded feature value LTP _P and the negative coded eigenvalue LTP _M are concatenated to form a feature dimension of 256×5×2, according to 2) in step (4) to calculate each The average precision of the method, the experimental results are shown in Table 1 and Figure 2. The coded images of several methods are shown in Fig. 3. In Fig. 3, the original picture, the 6 methods, and the coded images of this embodiment are given, wherein this embodiment is divided into a positive coded feature image LGDTP _P and a negative coded feature image LGDTP _M .

Table 1 Comparative experimental results between this embodiment and 6 methods (K=10)

serial number feature dimension Average precision (K=10) 1 LBP 1280 43.1%, 2 LDP 1280 45% 3 ELDP 1280 49.5%, 4 LDN 1280 47.3% 5 OLDP 1280 45.6%, 6 LTP 2560 52.1% 7 This embodiment 2560 63.3%

It can be seen from Table 1 and Figure 2 that when K is 10, the average precision of this embodiment is 63.3%. The local binary mode LBP method, the local direction mode LDP method, the enhanced local direction mode ELDP method, and the local direction number mode The average precision rates of the LDN method, the optimized local direction mode OLDP method, and the local ternary mode LTP method are 43.1%, 45%, 49.5%, 47.3%, 45.6%, and 52.1%, respectively. It can be seen that this method achieves the highest accuracy. average precision. The encoded image of the method of this embodiment is clearer.

Example 2

The images in this embodiment are from the tire track image database built by the applicant, including 30 categories of 80 images in each category, a total of 2,400 images. Experiments are carried out. The steps of the tire track image feature extraction method in the local gradient direction ternary mode are as follows:

(1) Image preprocessing

This procedure is the same as in Example 1.

(2) Feature extraction

1) Use the Sobel edge detection method to determine the image difference G _x along the x direction and the image difference G _y in the y direction, and use the formula (1) to determine the gradient direction angle α(x, y) of each pixel in the image

α(x,y)=arctan(G _y /G _x ) (13)

2) In the 3×3 neighborhood sliding window, use the local gradient direction ternary mode LGDTP method to determine each gradient direction angle value, and the local gradient direction ternary mode LGDTP method adds a custom threshold t, g _i is greater than the interval [g _c -t,g _c +t]} is 1, belongs to this interval is 0, is smaller than this interval is -1, and obtains a three-valued encoded value:

where P is the number of pixels in the neighborhood, R is the radius of the neighborhood, t in this embodiment is π/18, g _c is the gradient direction angle of the central pixel, and _gi is the gradient direction angle of its neighborhood pixels. The other steps in this step are the same as in Example 1.

Other steps are the same as in Example 1.

Example 3

The images in this embodiment are from the tire track image database built by the applicant, including 30 categories of 80 images in each category, a total of 2,400 images. Experiments are carried out. The steps of the tire track image feature extraction method in the local gradient direction ternary mode are as follows:

(1) Image preprocessing

This procedure is the same as in Example 1.

(2) Feature extraction

1) Use the Sobel edge detection method to determine the image difference G _x along the x direction and the image difference G _y in the y direction, and use the formula (1) to determine the gradient direction angle α(x, y) of each pixel in the image

α(x,y)=arctan(G _y /G _x ) (16)

2) In the 3×3 neighborhood sliding window, use the local gradient direction ternary mode LGDTP method to determine each gradient direction angle value, and the local gradient direction ternary mode LGDTP method adds a custom threshold t, g _i is greater than the interval [g _c -t,g _c +t]} is 1, belongs to this interval is 0, is less than this interval is -1, and obtains a three-value code value:

where P is the number of pixels in the neighborhood, R is the radius of the neighborhood, t in this embodiment is 3π/2, g _c is the gradient direction angle of the center pixel, and _gi is the gradient direction angle of its neighborhood pixels. The other steps in this step are the same as in Example 1.

Other steps are the same as in Example 1.

Example 4

In the image preprocessing step (1) of the above embodiment 1-3, 2400 tire track sample images of 30 categories and 80 images of each category are selected from the tire track image database, and the size is normalized to 768×768, gray Scale processing.

Step 1) of the step (3) of determining the feature vector in this embodiment is: dividing the positive coded image composed of the positive coded feature value LGDTP _P and the negative coded image composed of the negative coded feature value LGDTP _M into 3×3 subsections uniformly, respectively. The blocks are numbered 1 to 9 sub-regions from top to bottom in order, and the pixels of each sub-region are m×n, and m and n in this embodiment are 256. The other steps in this step are the same as in Example 1.

Other steps are the same as in Example 1.

Claims (4)

1. a tire trace image feature extraction method of a local gradient direction ternary pattern is characterized in that being made up of the following steps: (1) Image preprocessing Select 30 types of tire trace sample images from the tire trace image database, 50 to 80 for each type, and perform size normalization and grayscale processing; (2) Feature extraction 1) Use the Sobel edge detection method to determine the image difference G _x along the x direction and the image difference G _y in the y direction, and use the formula (1) to determine the gradient direction angle α(x, y) of each pixel in the image α(x,y)=arctan(G _y /G _x ) (1) 2) In the 3×3 neighborhood sliding window, use the local gradient direction ternary mode LGDTP method to determine each gradient direction angle value, and the local gradient direction ternary mode LGDTP method increases the custom threshold t, g _i is greater than the interval [g _c -t,g _c +t]} is 1, belongs to this interval is 0, is smaller than this interval is -1, and obtains a three-valued encoded value: where P is the number of neighborhood pixels, R is the radius of the neighborhood, 0 < t < 2π, g _c is the gradient direction angle of the central pixel, and _gi is the gradient direction angle of its neighborhood pixels; 3) Decompose the LGDTP eigenvalue into positive and negative coding eigenvalues, modify the coding value to 0 if the coding value is not 1, obtain the positive coding eigenvalue LGDTP _P , modify the coding value -1 to 1, and modify the remaining coding values to 0 to obtain the negative coding eigenvalue LGDTP _M ; 4) positive coding feature value LGDTP _P replaces the pixel value in the image to construct a positive coding image, and negative coding feature value LGDTP _M replaces the pixel value in the image to construct a negative coding image; (3) Determine the feature vector 1) Divide the positive coded image composed of the positive coding eigenvalue LGDTP _P and the negative coded image composed of the negative coding eigenvalue LGDTP _M into 3×3 sub-blocks uniformly, from top to bottom in order, and labelled 1 from left to right ~9 sub-regions, the pixels of each sub-region are m×n, and m and n are 128 or 256; 2) Perform histogram statistics on the gradient direction angles of the pixels in each sub-block of the positive-coded image, and cascade all the sub-block histograms; perform histogram statistics on the gradient direction angles of the pixels in each sub-block in the negative-coded image, level Concatenate all the sub-block histograms, and concatenate the sub-block histograms of the positive coded image and the negative coded image; 3) The concatenated histogram is used as the feature of the entire tire track image, the length of the features of the positive encoding image LGDTP _M and the negative encoding image LGDTP _P are 256 respectively, and the feature length of the final encoded image LGDTP is 256 × 2 × 5, Denoted as F _LGDTP ; 4) Normalize the F _LGDTP feature according to formula (4) to obtain the normalized eigenvector value: where F ^c (t) is the t-th component of feature c, is the normalized eigenvector value; (4) Image retrieval 1) Determine the eigenvector value of each tire track image Take each image as a query image and each image in the tire track image database to measure the similarity using the Manhattan distance metric method according to (5): where d is the distance between the features of the two tire track images, X _i , X _j represent the feature vector of each tire track image; 2) Using the average precision rate P as the retrieval performance evaluation index, it is determined by the following formula (6): Where S is the number of correct images contained in the query result, K is the total number of images in the query result, and K is less than the number of sheets of each category in the tire track image database. 2. The tire trace image feature extraction method of local gradient direction ternary mode according to claim 1, is characterized in that: in step 1) of determining feature vector step (3), described m and n are divisible 3 is a positive integer and is equal. 3. The tire trace image feature extraction method of the local gradient direction ternary pattern according to claim 1, is characterized in that: in step 2) of feature extraction step (2), t in formula (3) is π/ 6. 4. The tire track image feature extraction method of the local gradient direction ternary pattern according to claim 1, is characterized in that: in the step 2) of determining the feature vector step (3), according to the spatial distribution of tire track image texture information The particularity, the sub-block histogram is the histogram of sub-regions labeled 2, 4, 5, 6, and 8 respectively.