WO2021082168A1 - Method for matching specific target object in scene image - Google Patents

Abstract

A method for matching a specific target object in a scene image, relating to the field of image processing. For the problems, in the existing image clustering and feature matching algorithms, of high calculation complexity, low efficiency, and being difficult to adapt to the current big-data environment, the following method is proposed. The method comprises: performing superpixel image block segmentation on a scene image, extracting a central attribute of superpixels, calculating an adjacent matrix reflecting an adjacent relationship between superpixel image blocks, calculating a similarity matrix reflecting a superpixel between adjacent superpixelimage blocks, clustering superpixel image blocks according to the similarity matrix, performing image selection and feature extraction on a specific target object in the scene image, and searching for image blocks having a color close to a feature value of the target object. Said method improves the speed of matching calculation for a specific target object in a scene image, achieves efficient, convenient, and fast matching, and optimizes use of the technology in the fields of visual navigation, target measurement, and target tracking and positioning.

Description

A matching method for specific target objects in scene images Technical field

The invention belongs to the field of image processing, and relates to an application method based on an image clustering algorithm, in particular to an application method for specific target objects in scene images that can be used in the fields of visual navigation, target tracking and positioning, panoramic fusion, and three-dimensional simulation. Matching method.

Background technique

Image clustering is the use of a computer to analyze the images in the image library, and classify each pixel or area in the image into one of several feature categories to replace the human visual judgment of the image. The process of image clustering is essentially the process of image comprehension based on knowledge, and it is also the extension and development of humans' visual discrimination of images.

Image clustering technology is to search based on the semantic and perceptual characteristics of the image. The specific implementation is to extract specific information clues or feature indicators from the image data, and then search based on these clues from a large number of images stored in the image database. Image data with similar characteristics. Image clustering technology first clusters images according to a certain similarity principle, and aggregates similar images into one category. The retrieval process is carried out within the category, thereby greatly reducing the scope of image retrieval and achieving rapid and accurate image retrieval. purpose.

Image clustering technology has broad application prospects in all walks of life. For example, in the public security industry, with the continuous development of public security informatization, image recognition technology has been widely used in the public security industry. Video pictures have been obtained by means of camera capture and picture structuring, forming a dynamic resource library. Machine vision analysis technology based on image clustering can provide strong support for public security prevention and control, criminal investigation and solving, anti-terrorism and riot prevention. For another example, in the field of navigation, there is currently a visual automatic navigation system that uses the surrounding environment information to navigate through a camera installed on the vehicle body. The image information obtained by the camera can be analyzed and processed to obtain the position and posture information of the vehicle relative to the road, make corresponding path planning, and realize the automatic navigation of the vehicle.

The current conventional image clustering method is spectral clustering. The main advantage of the spectral clustering method is that spectral clustering only needs the similarity matrix between data, so it is very effective for processing sparse data clustering, which is difficult for traditional clustering algorithms such as K-Means. And because of the use of dimensionality reduction, the complexity of processing high-dimensional data clustering is better than traditional clustering algorithms. However, the main disadvantage of the spectral clustering method is that if the dimensionality of the final cluster is very high, the computational complexity of dimensionality reduction will be high, so the running speed of spectral clustering is slow and the final clustering effect is not ideal.

Feature matching is the matching of key target objects with the same or similar features in multiple images. It is a key link in image clustering and machine vision recognition technologies. It has important applications in the fields of panoramic fusion, monitoring, live broadcasting, and 3D simulation. The visual features adopted by current image clustering methods lack autonomous learning capabilities, resulting in poor image expression, high computational complexity, low clustering efficiency, and difficulty in adapting to the current big data environment. Therefore, it is difficult to adapt to the specific target in the scene image. The matching of objects is also very inefficient.

Therefore, finding an efficient and convenient image clustering method, and then realizing the matching of specific target objects in the scene image based on the image clustering method, has become an important foundation and an indispensable important link for image processing.

Summary of the invention

In view of the current image clustering and feature matching mentioned in the background art, the computational complexity is relatively high, the efficiency is low, and it is difficult to adapt to the current big data environment. The present invention proposes a matching of a specific target object in a scene image. The purpose of the method is to achieve high efficiency and convenience in matching specific target objects in the scene image.

In order to achieve the above objective, the present invention provides the following technical solutions:

A method for matching a specific target object in a scene image includes the following steps:

The first step is to segment the scene image into super pixel blocks and extract the super pixel center attributes in each super pixel block, where the super pixel center attributes include a position center and a color center;

The second step is to obtain an adjacency matrix reflecting the adjacency relationship between each super pixel block;

The third step is to obtain a similarity matrix reflecting the similarity between adjacent super-pixel tiles according to the adjacency matrix, where the similarity includes the similarity of the position adjacent relationship and the color;

The fourth step is to complete clustering of super pixel tiles according to the similarity matrix;

The fifth step is to perform image selection and feature extraction on specific target objects in the clustered scene images;

The sixth step is to search for color tiles that are similar to the feature value of the target object in the scene image.

Preferably, the superpixel center attribute includes the following attributes: coordinate center(x,y) in the image, color_info(l,a,b), superpixel unique identifier id labels, and number of superpixels num_pixels.

Preferably, the specific algorithm for calculating the adjacency matrix is implemented as follows:

Among them, i and j respectively represent the sequence numbers of the super-pixel tiles;

Each element e(i,j) in the adjacency matrix E satisfies the following functional relationship:

Among them, the relationship between the super pixel block itself and itself is defined as adjacent.

Preferably, the step of calculating the similarity matrix is to calculate the similarity of the two superpixels according to the adjacent relationship of the superpixel tiles in the adjacency matrix, when the similarity must be greater than a certain threshold, the corresponding element value is set to 1, otherwise it is set to 0, the specific algorithm is implemented as follows:

(1) Convert from CIE Lab color space to LθM color space

θ′=atan2(B,A) θ′∈(-π,π] (Formula 3-1)

l=L l∈[0, 100]

(2) Similarity calculation

Among them, L _th , θ _th , M _th , L _th0 , and θ _th0 are the threshold values of the three components in the LθM color space, and M _Cth is the threshold value for distinguishing color and black and white color spaces by modulus length component, usually the value is less than or equal to 2. L _i , L _j , θ _i , θ _j , M _i , M _j are the mean values of super pixel tiles i and j in the LθM color space respectively; w(i,j) is expressed as two super pixel tiles The similarity of, where a value of 1 is similar, and a value of 0 is dissimilar.

Preferably, the step of clustering is to generate a similarity matrix W using similarity w(i, j), and W is the clustering relationship graph.

Preferably, the implementation of the specific algorithm for completing clustering based on the similarity matrix W includes the step of converting the similarity matrix W into a triangular matrix,

Similarity matrix

Triangular matrix, set all the lower left corners to zero,

Preferably, the implementation of the specific algorithm for completing clustering based on the similarity matrix W includes: the step of completing clustering,

Perform clustering algorithm on triangular matrix

first step:

Starting from the nth row and nth column of the matrix, search all the arrays with 1 on the nth column. If the array with 1 on the nth column has only the nth row, then a(n,n)=1, otherwise a(n,n )=0.

The formula is as follows:

in case

a(n,n)=0

Then these arrays are logically ORed for each column (1, 2, 3...n) in the "row descending order" order, and the result is assigned to the non-zero array with the smallest row number among the n columns [0,0,... …A(i _min ,i _min ),……,a(i _min ,n-1),a(i _min ,n)]; the non-zero item or calculation algorithm of the column is as follows:

a(i _min ,n)=a(i _min ,n)∪...∪a(n,n)

a(i _min ,n-1)=a(i _min ,n-1)∪...∪a(n,n-1)

……………………………………

a(i _min ,i _min )=a(i _min ,i _min )∪...∪a(n,i _min )

Assignment operation:

a(n,n)=0

The end of this calculation;

The second step:

Starting from the n-1th row and n-1 column of the matrix, search all the arrays with 1 in the n-1 column. If the array with 1 in the n-1th column has only the n-1th row, then a(n-1 , N-1)=1 otherwise a(n-1, n-1)=0

The formula is as follows:

in case

a(n-1, n-1)=0

Then these arrays are logically ORed on each column (1, 2, 3...n-1) in the "row descending order" order, and the result is assigned to the non-zero array with the smallest row number in the n-1 column [0 ,0,……a(i _min ,i _min ),……,a(i _min ,n-1),a(i _min ,n)]; the non-zero items or calculation algorithms of the columns are as follows:

a(i _min ,j _n )=a(i _min ,j _n )∪...∪a(n,j _n )

a(i _min ,j _n-1 )=a(i _min ,j _n-1 )∪...∪a(n,j _n-1 )

……………………………………

a(i _min ,i _min )=a(i _min ,j _min )∪...∪a(n,i _min )

Assignment operation:

a(n-1, n～n-1)=0

, This calculation is over;

third step:

By analogy, starting from the i-th row and i-column of the matrix, search for all arrays with 1 on the i-th column. If the array with 1 on the i-th column has only the i-th row, then a(i, i) = 1 otherwise a( i, i) = 0

The formula is as follows:

in case

a(i,i)=0

Then these arrays are logically ORed for each column (1, 2, 3...n) in the "row descending order" order, and the result is assigned to the non-zero array with the smallest row number in column i [0,0,... …A(i _min ,i _min ),……,a(i _min ,n-1),a(i _min ,n)]; the non-zero term or calculation algorithm in the column is as follows:

a(i _min ,j _n )=a(i _min ,j _n )∪...∪a(n,j _n )

a(i _min ,j _n-1 )=a(i _min ,j _n-1 )∪...∪a(n,j _n-1 )

……………………………………

a(i _min ,i _min )=a(i _min ,j _min )∪...∪a(n,i _min )

Assignment operation:

a(i, n～n-i)=0

The end of this calculation;

the fourth step:

According to the above algorithm, each row of the triangular matrix is traversed once, and the following similar matrix will be obtained:

Then all the non-zero row arrays in the matrix are the arrays of clustering tiles.

Preferably, the image selection and feature extraction of the specific target object in the scene image refers to selecting the image block of the specific target object in the scene according to the image clustering result and extracting the values of L, θ, and M in the corresponding block.

Preferably, the specific method of searching for a color tile similar to the feature value of the target object tile is as follows:

Among them, L _th , θ _th , and M _th are the threshold values of the three components in the LθM color space, and M _Cth is the threshold value for distinguishing between color and black and white color spaces by modulus length components. Usually the value is less than or equal to 2, _Li , L _j ,θ _i ,θ _j ,M _i ,M _j are the mean values of super pixel tiles i and j in the LθM color space, respectively; where i represents the number of the selected target object in the sample image, and j represents the search image The block number of; w(i,j) represents the similarity of two superpixel blocks, where a value of 1 is similar, and a value of 0 is dissimilar;

Perform the following operations on the tiles of the search image:

If w(i,j)=1, keep the original value of the pixel in the search image block unchanged;

If w(i,j)=0, then the value of the pixel in the search image block is set to a value that is not in the color space, such as -1, then the block will not participate in the subsequent calculations;

Construct an adjacency matrix with the searched tiles:

Triangular matrix, set all the lower left corners to zero:

Use the same clustering method as the above to complete the clustering of the target object for the triangular matrix; thus achieve the matching of the target object in different scene images:

All non-zero row arrays in the matrix are the target object clustering tiles.

Due to the above scheme, the beneficial effects of the present invention are:

The image clustering method used in the present invention is a clustering method that simulates the process of human eye recognition of objects. First, the image is divided into super pixel blocks to extract the central attributes of the super pixels in each super pixel block, and then the reflection is calculated. The adjacency matrix of the adjacency between each super-pixel block, and then calculate the similarity matrix reflecting the similarity of the super-pixels between adjacent super-pixel blocks according to the adjacency matrix, and finally complete the super-pixel block according to the similarity matrix Clustering.

The calculation performance of the method of the present invention and the traditional spectral clustering and histogram clustering methods are as follows:

Computer configuration: CPU+GPU

Among them: CPU model i5=4590 main frequency: 3.3GHz;

The number of CUDA cores of the GPU is 2880 and the main frequency is 705MHz.

The resolution of the calculated image is 1920×1080

Comparison of computing performance of different image clustering methods:

Clustering method name	Number of iterations	Operation time (unit: second)
Spectral clustering	5	180
Histogram	5	60
New type spectrum clustering (this patent)	1	0.05

It can be seen from the above comparison table that the clustering method proposed by the present invention is significantly better than the traditional spectral clustering and histogram clustering methods in terms of operational performance.

The present invention proposes a matching method for a specific target object in a scene image on the basis of an image clustering method with a brand-new concept and based on the same concept. First, perform image selection and feature extraction on the specific target object in the clustered scene image; then search the scene image for color blocks that are similar to the feature value of the target object.

Through the technical scheme of the present invention, the calculation speed of the matching calculation of image clustering and specific target objects can be improved, thereby optimizing applications in the fields of target tracking and positioning, panoramic fusion, and three-dimensional simulation.

Description of the drawings

Figure 1 is the original image of the sample scene.

Figure 2 is an image after the sample cluster is segmented.

Figure 3 is the extraction of the corresponding tiles in the sample scene where the target objects are people and bags.

Detailed ways

The specific embodiments of the present invention will be described in more detail below in conjunction with the schematic diagrams. Based on the following description, the advantages and features of the present invention will become clearer. It should be noted that the drawings all adopt a very simplified form and all use imprecise proportions, which are only used to conveniently and clearly assist in explaining the purpose of the embodiments of the present invention.

In the following, an embodiment is used to verify the calculation process and the clustering and matching effects of the present invention.

Please refer to Figure 1, which is the original image of the sample scene.

The specific steps of the algorithm of the present invention are as follows:

Step 1. Recalculate the cluster center Seed.

This step is a process of labeling each pixel in the image, which makes the pixels with the same label have a certain common visual characteristic. The result of superpixel segmentation is a collection of sub-regions on the image. The entirety of these sub-regions covers the entire image, or a collection of contour lines extracted from the image, such as edge detection. Each pixel in a super pixel block is similar under a certain characteristic measurement or calculated characteristic, such as color, brightness, and texture. Adjacent regions are very different under a certain characteristic measurement.

In the field of computer vision, superpixels are widely used in the initial stage of image segmentation and understanding. The use of superpixels can effectively reduce the redundancy of image local information and reduce the complexity of image processing. Pixels are not the focus of human vision. Because humans obtain images from a region where many pixels are combined, a single pixel has no practical meaning, and only when combined together is meaningful to humans. So in this case there is the concept of "super pixels". The so-called super pixel is a small area in the image composed of a series of adjacent pixels with similar characteristics such as color, brightness, texture, etc. Most of these small areas retain effective information for further image segmentation, and generally will not damage Boundary information of objects in the image. Therefore, substituting superpixels for the original pixels as nodes of the graph for image segmentation can greatly reduce the scale of image processing and bring computational advantages.

In the present invention, the central attributes of superpixels are defined as follows:

The code is implemented as follows:

The above code is for reference only.

Step 2: Calculate the adjacency matrix E.

This step of the present invention takes into account that since in the clustering of super pixel blocks, only the mutual clustering between adjacent super pixel blocks needs to be considered, and no calculation is required for non-adjacent super pixel blocks. So we first give the adjacency matrix E, which is used to calculate the adjacency matrix for the subsequent similarity clustering.

The present invention uses parallel computing:

(Note: i and j respectively represent the serial numbers of super pixel tiles)

Each element e(i,j) in the adjacency matrix E satisfies the following functional relationship:

(Note: The relationship between the super pixel block itself and itself is defined as adjacent)

Step three, similarity matrix W

Similarity measure is a function used to compare images. The similarity between images or parts of images is a very important issue at the bottom of the computer vision field. For the image clustering algorithm we proposed, similarity plays a decisive and key role, and different similarity measures will lead to completely different clustering effects.

The algorithm idea of this step of the present invention is to calculate the similarity of two superpixels based on the adjacent relationship of the superpixel tiles in the adjacency matrix E. When the similarity must be greater than a certain threshold, the corresponding element value is set to 1, otherwise it is set to 0. The specific algorithm is implemented as follows.

(Note: The algorithm can be changed to different parameters and formulas according to different scenarios)

Calculated as follows:

The first is the transformation of color space, CIE

That is, convert from CIE Lab space to LθM space.

This step effectively simulates the conversion of human recognition methods based on the surface color and brightness of the object under different color saturation conditions, and realizes the effective clustering of objects with different color saturation in the scene image, which improves the clustering of the image. Effectiveness and anti-interference ability, the dimensionality reduction effect of image clustering and segmentation is obvious, which can effectively improve the efficiency and accuracy of image analysis.

For details of this color space, please refer to the color space described in the Chinese patent application document "Color image clustering and segmentation method based on multi-scale perception characteristics of human vision" with the applicant's publication number CN104063707A and patent number ZL201410334974.3.

θ′=atan2(B,A) θ′∈(-π,π] (Formula 3-1)

l=L l∈[0, 100]

The code is implemented as follows:

Then, the similarity calculation

Among them, L _th , θ _th , M _th , L _th0 , and θ _th0 are the threshold values of the three components in the LθM color space, and M _Cth is the threshold value for distinguishing color and black and white color spaces by the modulus component, usually the value is less than or equal to 2 , L _i , L _j , θ _i , θ _j , M _i , and M _j are the mean values of super pixel tiles i and j in the LθM color space, respectively. w(i,j) is expressed as the similarity of two superpixel tiles, where a value of 1 means similarity, and a value of 0 means dissimilar.

Step four, clustering

The algorithm of this step is to use the similarity w(i,j) to generate the similarity matrix W (W is the clustering relationship graph);

The algorithm steps to complete clustering based on the similarity matrix W are as follows:

First, the similarity matrix W is transformed into a triangular matrix

Similarity matrix

Triangular matrix (set all the lower left corners to zero)

Then, complete the clustering

Perform clustering algorithm on triangular matrix

first step:

Starting from the nth row and nth column of the matrix, search all the arrays with 1 on the nth column. If the array with 1 on the nth column has only the nth row, then a(n,n)=1, otherwise a(n,n )=0.

The formula is as follows:

in case

a(n,n)=0

Then these arrays are logically ORed for each column (1, 2, 3...n) in the "row descending order" order, and the result is assigned to the non-zero array with the smallest row number among the n columns [0,0,... …A(i _min ,i _min ),……,a(i _min ,n-1),a(i _min ,n)]. The non-zero term OR calculation algorithm of the column is as follows:

a(i _min ,n)=a(i _min ,n)∪...∪a(n,n)

a(i _min ,n-1)=a(i _min ,n-1)∪...∪a(n,n-1)

……………………………………

a(i _min ,i _min )=a(i _min ,i _min )∪...∪a(n,i _min )

Assignment operation:

a(n,n)=0

This calculation is over.

The second step:

Starting from the n-1th row and n-1 column of the matrix, search all the arrays with 1 in the n-1 column. If the array with 1 in the n-1th column has only the n-1th row, then a(n-1 , N-1)=1 otherwise a(n-1, n-1)=0

The formula is as follows:

in case

a(n-1, n-1)=0

Then these arrays are logically ORed on each column (1, 2, 3...n-1) in the "row descending order" order, and the result is assigned to the non-zero array with the smallest row number in the n-1 column [0 ,0,……a(i _min ,i _min ),……,a(i _min ,n-1),a(i _min ,n)]. The non-zero term OR calculation algorithm of the column is as follows:

a(i _min ,j _n )=a(i _min ,j _n )∪...∪a(n,j _n )

a(i _min ,j _n-1 )=a(i _min ,j _n-1 )∪...∪a(n,j _n-1 )

……………………………………

a(i _min ,i _min )=a(i _min ,j _min )∪...∪a(n,i _min )

Assignment operation:

a(n-1, n～n-1)=0

, This calculation is over.

third step:

By analogy, starting from the i-th row and i-column of the matrix, search for all arrays with 1 on the i-th column. If the array with 1 on the i-th column has only the i-th row, then a(i, i) = 1 otherwise a( i, i) = 0

The formula is as follows:

in case

a(i,i)=0

Then these arrays are logically ORed for each column (1, 2, 3...n) in the "row descending order" order, and the result is assigned to the non-zero array with the smallest row number in column i [0,0,... …A(i _min ,i _min ),……,a(i _min ,n-1),a(i _min ,n)]. The non-zero term OR calculation algorithm in the column is as follows:

a(i _min ,j _n )=a(i _min ,j _n )∪...∪a(n,j _n )

a(i _min ,j _n-1 )=a(i _min ,j _n-1 )∪...∪a(n,j _n-1 )

……………………………………

a(i _min ,i _min )=a(i _min ,j _min )∪...∪a(n,i _min )

Assignment operation:

a(i, n～n-i)=0

, This calculation is over.

the fourth step:

According to the above algorithm, each row of the triangular matrix is traversed once, and the following similar matrix will be obtained:

Then all the non-zero row arrays in the matrix are the arrays of clustering tiles.

Please refer to Figure 2 for the image after the sample clustering segmentation.

Step 5. Image selection and feature extraction of specific target objects in the scene

first step:

According to the result of image clustering, the image block of a specific target object is selected in the scene and the values of L, θ, and M in the corresponding block are extracted.

For example, the target objects in the scene shown in FIG. 3 are the L, θ, and M feature values of people and bags and corresponding tiles. From this, the feature values of the target object block in the image are as follows:

Character's tile feature value

Tile name	L	θ	M
head	52.9208	65.7256	8.59028
collar	50.6847	264.383	6.21815
coat	56.4843	256.406	8.63319
Arm 1	52.2053	80.9778	11.6405
Arm 2	54.0127	32.8274	3.96128
Arm 3	54.9561	75.5989	17.0923
Pants 1	42.6731	230.979	3.41989
Pants 2	37.0021	178.001	1.57844
Pants 3	48.9932	232.782	4.12733
Pants 4	47.4722	187.426	1.55389
Pants 5	43.5705	220.827	2.29207
Pants 6	48.948	241.587	3.6887
Pants 7	48.8031	200.716	1.69321
Shoes 1	55.207	210.237	3.21031
Shoes 2	52.885	181.793	3.26188

Block eigenvalues of the package

Tile name	L	θ	M
Pack of yellow blocks	78.3918	94.3349	49.0715
Pack of red blocks 1	83.085	20.7271	11.8367
Pack of red blocks 2	58.9622	27.184	34.9523

Note: The above colors refer to the actual colors of the sample images. The drawings in this application file only provide black and white drawings for reference due to the format regulations.

Step 6: Search for the target object

According to the image clustering method of steps 1 to 5 as described above, a scene clustering segmentation map is obtained, and (Equation 3-4) is used to search for color blocks with similar feature values of the target object blocks in the scene. The specific method is as follows:

Among them, L _th , θ _th , and M _th are the threshold values of the three components in the LθM color space, and M _Cth is the threshold value for distinguishing between color and black and white color spaces by modulus length components. Usually the value is less than or equal to 2, _Li , L _{j , θ i, θ j, M} i, M j respectively superpixel tile i, j Means LθM color space. Where i represents the block number of the selected target object in the sample image, and j represents the block number of the search image. w(i,j) is expressed as the similarity of two superpixel tiles, where a value of 1 means similarity, and a value of 0 means dissimilar.

Perform the following operations on the tiles of the search image:

If w(i,j)=1, keep the original value of the pixel in the search image block unchanged;

If w(i,j)=0, then the value of the pixel in the search image block is set to a value that is not in the color space, such as -1, and the block will not participate in the subsequent calculations.

Then, construct the adjacency matrix of the searched tiles:

Triangular matrix, set all the lower left corners to zero:

The clustering of the target object is also completed using the method of completing the clustering in step 4. So as to achieve the matching of target objects in different scene images.

All non-zero row arrays in the matrix are the target object clustering tiles.

The foregoing are only preferred embodiments of the present invention, and do not play any restrictive effect on the present invention. Any person skilled in the art, without departing from the scope of the technical solution of the present invention, makes any form of equivalent replacement or modification or other changes to the technical solution and technical content disclosed by the present invention, which does not depart from the technical solution of the present invention. The content still falls within the protection scope of the present invention.

Claims (9)

1. A method for matching a specific target object in a scene image is characterized in that it comprises the following steps:

   The first step is to segment the scene image into super pixel blocks and extract the super pixel center attributes in each super pixel block, where the super pixel center attributes include a position center and a color center;

   The second step is to obtain an adjacency matrix reflecting the adjacency relationship between each super pixel block;

   The third step is to obtain a similarity matrix reflecting the degree of similarity between adjacent super-pixel tiles according to the adjacency matrix, where the similarity includes the degree of similarity of the position adjacent relationship and the color;

   The fourth step is to complete clustering of super pixel tiles according to the similarity matrix;

   The fifth step is to perform block selection and feature value extraction for specific target objects in the clustered scene image;

   The sixth step is to search for color tiles that are similar to the feature value of the target object in the scene image.
2. The method for matching a specific target object in a scene image according to claim 1, wherein the superpixel center attribute includes the following attributes: the coordinate center (x, y) at the center of the image, the color average color_info (l, a ,b), the unique identification id labels of super pixels, the number of super pixels num_pixels.
3. The method for matching a specific target object in a scene image according to claim 2, wherein the specific algorithm for calculating the adjacency matrix is implemented as follows:

   

   Among them, i and j respectively represent the sequence numbers of the super-pixel tiles;

   Each element e(i,j) in the adjacency matrix E satisfies the following functional relationship:

   

   Among them, the relationship between the super pixel block itself and itself is defined as adjacent.
4. The method for matching a specific target object in a scene image according to claim 3, wherein the step of calculating the similarity matrix is to calculate the similarity of two adjacent super pixels according to the adjacent relationship of the super pixel blocks in the adjacency matrix. When the similarity must be greater than a certain threshold, the corresponding element value is set to 1, otherwise it is set to 0. The specific algorithm is implemented as follows:

   (1) Convert from CIE Lab color space to LθM color space

   θ′=atan2(B,A) θ′∈(-π,π] (Formula 3-1)

   

   

   l=L l∈[0, 100]

   (2) Similarity calculation

   

   Among them, L _th , θ _th , M _th , L _th0 , and θ _th0 are the threshold values of the three components in the LθM color space, and M _Cth is the threshold value for distinguishing color and black and white color spaces by modulus length component, usually the value is less than or equal to 2. L _i , L _j , θ _i , θ _j , M _i , M _j are the mean values of super pixel tiles i and j in the LθM color space respectively; w(i,j) is expressed as two super pixel tiles The similarity of, where a value of 1 is similar, and a value of 0 is dissimilar.
5. The method for matching a specific target object in a scene image according to claim 4, wherein the step of clustering is to generate a similarity matrix W using similarity w(i, j), and W is the clustering relationship graph .
6. The method for matching a specific target object in a scene image according to claim 5, wherein the implementation of the specific algorithm for completing clustering based on the similarity matrix W includes the step of converting the similarity matrix W into a triangular matrix,

   Similarity matrix

   

   Triangular matrix, set all the lower left corners to zero,

   
7. The method for matching a specific target object in a scene image according to claim 6, wherein the implementation of the specific algorithm for completing clustering based on the similarity matrix W includes: completing the step of clustering, performing clustering on the triangular matrix algorithm

   

   first step:

   Starting from the nth row and nth column of the matrix, search all the arrays with 1 on the nth column. If the array with 1 on the nth column has only the nth row, then a(n,n)=1, otherwise a(n,n )=0

   The formula is as follows:

   

   in case

   a(n,n)=0

   Then these arrays are logically ORed for each column (1, 2, 3...n) in the "row descending order" order, and the result is assigned to the non-zero array with the smallest row number among the n columns [0,0,... …A(i _min ,i _min ),……,a(i _min ,n-1),a(i _min ,n)]; the non-zero item or calculation algorithm of the column is as follows:

   a(i _min ,n)=a(i _min ,n)∪...∪a(n,n)

   a(i _min ,n-1)=a(i _min ,n-1)∪...∪a(n,n-1)

   ……………………………………

   a(i _min ,i _min )=a(i _min ,i _min )∪...∪a(n,i _min )

   Assignment operation:

   a(n,n)=0

   The end of this calculation;

   The second step:

   Starting from the n-1th row and n-1 column of the matrix, search all the arrays with 1 in the n-1 column. If the array with 1 in the n-1th column has only the n-1th row, then a(n-1 , N-1)=1 otherwise a(n-1, n-1)=0

   The formula is as follows:

   

   in case

   a(n-1, n-1)=0

   Then these arrays are logically ORed on each column (1, 2, 3...n-1) in the "row descending order" order, and the result is assigned to the non-zero array with the smallest row number in the n-1 column [0 ,0,……a(i _min ,i _min ),……,a(i _min ,n-1),a(i _min ,n)]; the non-zero items or calculation algorithms of the columns are as follows:

   a(i _min ,j _n )=a(i _min ,j _n )∪...∪a(n,j _n )

   a(i _min ,j _n-1 )=a(i _min ,j _n-1 )∪...∪a(n,j _n-1 )

   ……………………………………

   a(i _min ,i _min )=a(i _min ,j _min )∪...∪a(n,i _min )

   Assignment operation:

   a(n-1, n～n-1)=0

   The end of this calculation;

   third step:

   By analogy, starting from the i-th row and i-column of the matrix, search for all arrays with 1 on the i-th column. If the array with 1 on the i-th column has only the i-th row, then a(i, i) = 1 otherwise a( i, i) = 0

   The formula is as follows:

   

   in case

   a(i,i)=0

   Then these arrays are logically ORed for each column (1, 2, 3...n) in the "row descending order" order, and the result is assigned to the non-zero array with the smallest row number in column i [0,0,... …A(i _min ,i _min ),……,a(i _min ,n-1),a(i _min ,n)]; the non-zero term or calculation algorithm in the column is as follows:

   a(i _min ,j _n )=a(i _min ,j _n )∪...∪a(n,j _n )

   a(i _min ,j _n-1 )=a(i _min ,j _n-1 )∪...∪a(n,j _n-1 )

   ……………………………………

   a(i _min ,i _min )=a(i _min ,j _min )∪...∪a(n,i _min )

   Assignment operation:

   a(i, n～n-i)=0

   , This calculation is over;

   the fourth step:

   According to the above algorithm, each row of the triangular matrix is traversed once, and the following similar matrix will be obtained:

   

   Then all the non-zero row arrays in the matrix are the arrays of clustering tiles.
8. The method for matching a specific target object in a scene image according to claim 7, wherein the image selection and feature extraction of the specific target object in the scene image refers to the selection of the specific target object in the scene according to the result of image clustering. Image block and extract the values of L, θ, M in the corresponding block.
9. The method for matching a specific target object in a scene image according to claim 8, wherein the specific method for searching for a color tile similar to the feature value of the target object tile is as follows:

   

   Among them, L _th , θ _th , and M _th are the threshold values of the three components in the LθM color space, and M _Cth is the threshold value for distinguishing between color and black and white color spaces by modulus length components. Usually the value is less than or equal to 2, _Li , L _j ,θ _i ,θ _j ,M _i ,M _j are the mean values of super pixel tiles i and j in the LθM color space, respectively; where i represents the number of the selected target object in the sample image, and j represents the search image The number of the block; w(i,j) represents the similarity of two superpixel blocks, where a value of 1 is similar, and a value of 0 is dissimilar;

   Perform the following operations on the tiles of the search image:

   If w(i,j)=1, keep the original value of the pixel in the search image block unchanged;

   If w(i,j)=0, then the value of the pixel in the search image block is set to a value that is not in the color space, such as -1, then the block will not participate in the subsequent calculations;

   Construct an adjacency matrix with the searched tiles:

   

   Triangular matrix, set all the lower left corners to zero:

   

   Using the same clustering method as in claim 7 for the triangular matrix to complete the clustering of the target object; thereby achieving the matching of the target object in different scene images:

   

   All non-zero row arrays in the matrix are the target object clustering tiles.

Images