Salient object detection method based on heuristic boundary optimization

Abstract

The present invention belongs to the field of computer vision technology. In order to solve the problem that the saliency map obtained by the unsupervised saliency target detection method has obvious block phenomenon and poor edge details, the present invention is a saliency target detection method based on heuristic boundary optimization, and the steps are as follows: Step 1: obtain a coarse saliency map based on simple linear iterative clustering and basic digital image features; Step 2: obtain an image optimization mask by applying a region partitioning algorithm based on inter-class and intra-class variances; Step 3: construct a boundary optimization pyramid using a color image and the image optimization mask; Step 4: optimize the boundary area by applying a boundary optimization algorithm based on a dynamic population gray wolf algorithm, and obtain a saliency map with uniform target area and clear and complete boundaries. The present invention is mainly used in image processing occasions.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer vision, and in particular relates to a salient object detection method based on heuristic boundary optimization. Background Art

[0002] Salient object detection is a technology that uses computers to simulate the visual attention mechanism of the human eye. It can quickly segment the most eye-catching objects or areas from an image. Vision is one of the important ways for people to perceive and understand the world. With the development of science and technology, computer vision technology has attracted more and more attention. At present, salient object detection has become an important research topic in the field of computer vision, and as an image preprocessing technology, it has been widely used in other computer vision tasks, such as target recognition, target tracking, semantic segmentation, video compression, etc. Salient object detection can effectively guide the redundancy suppression of images and has important practical application value in the era of big data.

[0003] Early salient object detection methods mainly detected salient objects by designing manual features, such as texture, gradient, grayscale, etc. However, due to the limitations of the feature expression ability of manual features and the difficulty of effectively using features through basic digital image processing, it is difficult for salient object detection based on traditional methods to obtain a complete salient map of the target area. Deep neural networks have gradually attracted the attention of researchers due to their powerful feature extraction and learning capabilities, and have been successfully applied to the field of salient object detection. However, methods based on supervised neural networks have problems such as high computational cost, difficulty in labeling, and high dependence on data sets. Although the salient object detection method based on unsupervised neural networks can reduce the cost of labeling, due to insufficient supervision information, the network cannot detect salient maps with fine boundaries, which makes it difficult to meet the accuracy requirements of practical applications. Summary of the invention

[0004] In order to overcome the shortcomings of the prior art, the present invention aims to mainly solve the problems of obvious blocking phenomenon and poor edge details of the saliency map obtained by the unsupervised saliency target detection method, and proposes a saliency target detection method based on heuristic boundary optimization.

[0005] To this end, the technical solution adopted by the present invention is a salient object detection method based on heuristic boundary optimization, the steps are as follows:

[0006] Step 1: Obtain a coarse saliency map based on simple linear iterative clustering and basic digital image features;

[0007] Step 2: Apply the region partitioning algorithm based on inter-class and intra-class variance to obtain the image optimization mask, perform region partition on the coarse saliency map obtained in step 1, and obtain the image optimization mask having three parts: the main area, the area to be optimized, and the background area;

[0008] Step 3: Use the color image and the image optimization mask to construct a boundary optimization pyramid. By performing sampling operations with the same parameters on the color image and the image optimization mask at the same time, a multi-scale representation of the image, namely, a boundary optimization pyramid, is constructed.

[0009] Step 4: Apply the boundary optimization algorithm based on the dynamic population gray wolf algorithm to optimize the boundary area. The dynamic population-based gray wolf optimization algorithm sets the population dynamic rate and outlier rate parameters on the basis of the original gray wolf algorithm, and uses the new search equation to improve the global optimization ability and multi-peak function optimization ability of the algorithm. By refining the boundary of the image optimization mask, a saliency map with uniform target area and clear and complete boundaries is obtained.

[0010] The specific steps are as follows:

[0011] Step 1: Based on simple linear iterative clustering and basic digital image features, a coarse saliency map is obtained. The simple linear iterative clustering SLIC method is used to divide the image into multiple regions, namely superpixel regions. In the superpixel region, pixels are grouped or clustered according to human visual perception. Multiple homogeneous regions can be extracted in an image. Then, the feature contrast map is extracted by measuring the color contrast and spatial contrast between superpixel regions. In addition, the input image is subjected to corner detection by the Harris operator to obtain the salient points in the input image. On this basis, the spatial distribution of superpixel regions is used to highlight the salient target and suppress the background. Then, the probability of each superpixel region being the background is determined, and the background connectivity is represented by a numerical value. The feature contrast map, spatial compactness term, and background correlation term are smoothed respectively and fused together to generate an initial salient feature map. Then, the cost function is used to fuse the initial salient feature map and the background probability map to obtain a coarse saliency map.

[0012] Step 2: Apply the regional division algorithm based on inter-class and intra-class variance to obtain the image optimization mask: Apply the regional division algorithm based on inter-class and intra-class variance to divide the coarse saliency map obtained in step 1 into three parts: the main area, the area to be optimized, and the background area. Use the criteria of maximum inter-class variance and minimum intra-class variance to construct the cost function, calculate the image optimization mask segmentation thresholds T1 and T2, segment the coarse saliency prediction map based on T1 and T2, and remove noise through opening operation to obtain the image optimization mask;

[0013] Step 3: Use the color image and the image optimization mask to construct a boundary optimization pyramid. At the same time, perform sampling operations with the same parameters on the color image and the image optimization mask to construct a multi-scale representation of the image, namely, the boundary optimization pyramid. To avoid the boundary optimization pyramid from containing too small an image, a boundary optimization pyramid establishment stop criterion is set. When the boundary optimization pyramid establishment stop criterion is met, the algorithm stops calculating and obtains a boundary optimization pyramid data model.

[0014] Step 4: Apply the boundary optimization algorithm based on the dynamic population gray wolf algorithm to optimize the boundary area. Before the iterative search, the gray wolf optimization algorithm based on the dynamic population is used for optimization. The algorithm provides the initial pixel pairs for heuristic information propagation. For the boundary optimization pyramid data model, starting from the smallest scale image, multi-scale pixel pair search for heuristic information propagation is performed step by step, and the heuristic information obtained from the previous layer of the pyramid is propagated to the current layer, and heuristic information propagation and single-scale search are performed alternately; Among them, the dynamic population gray wolf optimization algorithm sets the population dynamic rate and outlier rate parameters on the basis of the original gray wolf algorithm, and uses the new search equation to improve the algorithm's global optimization ability and multi-peak function optimization ability. By refining the boundary of the image optimization mask, a saliency map with uniform target area and clear and complete boundaries is obtained.

[0015] The detailed steps are as follows:

[0016] Step 1: Obtain a coarse saliency map based on simple linear iterative clustering and basic digital image features: Use the simple linear iterative clustering SLIC (Simple Linear Iterative Cluster) method to segment the image into multiple regions, namely superpixel regions. Within the superpixel region, pixels are grouped or clustered according to human visual perception; quantize each color channel of the input image and quantize it uniformly into q binary numbers. From the quantized image, select the main color to cover more than 95% of the pixels in the image. According to the minimum Euclidean distance criterion, the color covering less than 5% of the pixels is merged into the selected main color. After image segmentation and color quantization based on superpixels, each superpixel region S i Each has its histogram H i , i = 1, 2, ... n, and Normalize, where k is the color quantization number; then, extract the feature contrast map by measuring the color contrast and spatial contrast between superpixel regions. i and S j , their color distance C d (S i ,S j ) is defined as the Euclidean distance between the average colors of two superpixel regions, ranging from [0,1]. Here, the color distance Cd Evaluate the color similarity of two superpixel regions. In addition to the color-based similarity, i and S j The Gestalt law of proximity is expressed as the spatial proximity distance P d (S I ,S j ), which is defined as the Euclidean distance between the centroids of two superpixels, P d (S I ,S j ) is larger, S i and S j The closer the distance is, the closer the spatial distance can be further normalized in [0,1] as The formula is:

[0017]

[0018] Where P d-min and P d-max P d The minimum and maximum of all possible values, determined by the superpixels of a given image;

[0019] Based on color distance C d and standardized spatial proximity The global superpixel color contrast of Si is obtained, which is defined as:

[0020]

[0021] Where n is the number of superpixel regions, A j For super pixel S j The normalized area of ​​​​the large S i The global color contrast value of contributes more, and the sum of Aj is 1;

[0022] Define the superpixel region S i The global spatial contrast values ​​are as follows:

[0023]

[0024] Where L j is the weight of the spatial layout, which is defined as the superpixel region S j The minimum distance from the center of mass of the image to any of the four borders of the image; C I is the color similarity between superpixel regions defined by equation (4), where Hi and Hj refer to two superpixel regions S i and S j The color histogram of , m is the index of the color quantization number k;

[0025]

[0026] In the formula, by (1-|H i (m)-H j (m)|) measures the similarity of two histograms based on frequency; ICS is defined as the superpixel region S i and S j A combined measure of color similarity and spatial proximity between:

[0027] ICS(S i ,S j )=P d-norm (S i ,S j )·C I (S i ,S j ) (5)

[0028] Therefore, the formula for the global spatial contrast value can be simplified to:

[0029]

[0030] In the global color contrast C G and the global spatial contrast S G Based on this, we get the super pixel area S i The characteristic contrast value FC(S i ), whose formula is:

[0031] FC(S i )=C G (S i )·S G (S i ) (7)

[0032] In addition, the input image is subjected to corner detection through the Harris operator to obtain the salient points in the input image. On this basis, the spatial distribution of the superpixel area is used to highlight the salient target and suppress the background.

[0033] For a given superpixel region image, the colors of all superpixel regions can also be divided into 8 coarse clusters in the Lab color space, where the 3 individual color components are separated by the threshold T (.) Divided into two parts, for each of the 8 color clusters, the extracted main color is recorded as in is the average value of one of the three color components of Lab, C i As the center of the corresponding color cluster, the Harris operator is used to determine the coarse area of ​​the foreground and background, and the convex hull is calculated to surround all salient points. All superpixel areas within the convex hull are defined as the foreground, and the remaining superpixels are defined as the background. and Represent the average color of the foreground Gf and background Gb respectively. In order to maximize the color contrast, in each color component, the adaptive threshold T (.) is defined as:

[0034]

[0035] According to the 8 main colors adaptively determined above, similar colors are merged to simplify subsequent calculations. Each pair of extracted main colors is compared. If their Euclidean distance is less than the threshold T d , then merge them, and for each image, adaptively determine the threshold as and Half of the Euclidean distance between them, let Ci and Cj be two color clusters, and merge them using the weighted average agglomeration method:

[0036]

[0037] Where pi and pj represent C i and C j The number of pixels in G f or G b Each color cluster C in k ,M(C k ) represents its geographical center, C k The spatial compactness term is defined as M(C k ) with each object O i The average distance between , which contains:

[0038]

[0039]

[0040] Where k≤8, N k Represents color cluster C k The number of objects in M(O i ) represents the geographic center of the object O i . Through formula (12), Sd is scaled to [0,1];

[0041]

[0042] Where Sd(C k ) is Sd f (C k ) or Sd b (C k );Sd min and Sd max are the minimum and maximum values ​​of all spatial compactness Sd, respectively.

[0043] The background correlation term is calculated to measure the degree of connection between object O and the image boundary. The larger the value, the higher the overlap between O and the image boundary. The background correlation term is defined as:

[0044]

[0045] Where N c is the total number of superpixels in object O, N b is the number of superpixels in object O that are directly connected to the image boundary, and B c The maximum and minimum values ​​of all possible values ​​of are denoted by B c_max and B c_min , get the normalized B in the range [0,1] c , the formula is:

[0046]

[0047] After that, the probability of each superpixel region being the background is determined, and the background connectivity is represented by a numerical value. For a given superpixel region S i , the probability that it is the background can be expressed by formula (15):

[0048]

[0049]

[0050]

[0051] Where σ is set to 10, δ(·) is 1 when the superpixel region is on the image boundary, otherwise δ(·) is 0, n is the number of superpixel regions, and D geo (S i ,S j ) is the geodesic distance between any two superpixel regions, which is determined by the cumulative weight d app It is obtained by following the adjacent superpixel area S in the figure i and S j The Euclidean distance of the shortest path between sequences is measured, and l is the total number of superpixel regions in the determined path;

[0052]

[0053] In order to make BP in [0,1], σ BP Set to 1, the normalization process is defined as:

[0054]

[0055] The feature contrast map, spatial compactness term, and background correlation term are smoothed separately and fused together to generate the initial salient feature map. For each superpixel, its saliency value is replaced by the weighted average of the saliency values ​​of the surrounding superpixel regions. For a given superpixel region S i , select The nearest neighbor is weighted averaged and the local refinement S i The characteristic contrast value is as follows:

[0056]

[0057] in It is in S i The sum of the color similarity and spatial similarity between it and its k nearest neighboring regions;

[0058] For C k All superpixel regions in will be assigned the same spatial compactness term N Sd (C k ), the lower this value is, the higher its similarity with the salient object is. After that, the spatial compactness term of each superpixel region Si is smoothed, and its formula is as follows:

[0059]

[0060] Where n is the number of superpixel regions, YesS i The sum of color similarity and spatial proximity to other superpixel regions;

[0061] In addition, the background related term is calculated by formula (22):

[0062]

[0063] Then, the space compactness term R Sd (S i ) and background related items Combined, determine the background suppression map:

[0064]

[0065] The initial saliency feature map is expressed as:

[0066] ISA(S i )=RFC(S i )·OP(S i ) (twenty four)

[0067] Next, the cost function is used to fuse the initial saliency feature map and the background probability map to obtain a rough saliency map. Let FSA i is the super pixel area Si The final significance value of

[0068]

[0069]

[0070] Step 2: Assume that the image thresholds are T1 and T2, and the global average is m g , pixels less than T1 are background, with an average gray value of u0; pixels greater than T1 and less than T2 are areas to be optimized, with an average gray value of u1; pixels greater than T2 are target areas, with an average gray value of u2. The two thresholds are obtained using the criteria of maximum inter-class variance and minimum intra-class variance. The inter-class variance calculation formula is:

[0071] g inter =ω0×(u0-m g ) 2 +ω1×(u1-m g ) 2 +ω2×(u2-m g ) 2 (26)

[0072] Among them, ω0=3, ω1=1, ω2=3 are the weight coefficients of the three regions respectively. Since it is easier to obtain more accurate reasoning in the main area and the background area, the weight coefficients corresponding to the two areas are larger, m g is the global average gray value; the intra-class variance calculation formula is:

[0073]

[0074] Among them, m0, m1, and m2 are the number of pixels in the main area, the area to be optimized, and the background area respectively. i , x j , x k Then the pixel values ​​of the three regions are respectively;

[0075] The threshold segmentation cost function is constructed based on the inter-class variance and intra-class variance, and its formula is:

[0076] g=α g ×g inter ×β g -(1-α g )×g intra (28)

[0077] Among them, α g represents the inter-class variance ratio coefficient, α g =0.5,β gRepresents the inter-class variance weight coefficient, adjusts the values ​​of T1 and T2, and when g reaches the maximum value, T1 and T2 are the optimal thresholds. The rough saliency prediction map is segmented based on T1 and T2, and the noise is removed through the opening operation to obtain the image optimization mask;

[0078] Step 3: Use the color image and the image optimization mask to construct a boundary optimization pyramid. At the same time, perform sampling operations with the same parameters on the color image and the image optimization mask to construct a multi-scale representation of the image, namely, the boundary optimization pyramid. To avoid the boundary optimization pyramid from containing too small images, set a boundary optimization pyramid establishment stop criterion: when one of the length or width of the image drops below 100 pixels, the downsampling loop terminates. When the boundary optimization pyramid establishment stop criterion is met, the algorithm stops calculating and obtains the boundary optimization pyramid data model.

[0079] Step 4: Apply the boundary optimization algorithm based on the dynamic population gray wolf algorithm to optimize the boundary area. Before the iterative search, the gray wolf optimization algorithm based on the dynamic population is used to optimize, which provides the initial pixel pairs for heuristic information propagation. For the boundary optimization pyramid data model, starting from the smallest scale image, multi-scale pixel pair search for heuristic information propagation is performed step by step, and the heuristic information obtained from the previous layer of the pyramid is propagated to the current layer, and heuristic information propagation and single-scale search are performed alternately;

[0080] In the heuristic information propagation, the boundary optimization pyramid is used to define the neighborhood of pixel pairs on adjacent levels. The pixels in the current layer of the pyramid used to calculate the average color of the pixels in the previous layer are defined as the neighborhood of the pixels in the previous layer. The neighborhood relationship of pixels of different scales can be extended to pixel pairs. In the process of selecting pixel pairs, the optimization algorithm is not applied to high-quality pixel pairs. Only half of the low-quality pixel pairs are selected from the initial decision vector to be optimized to avoid redundant search.

[0081] The gray wolf optimization algorithm for dynamic populations sets the population dynamic rate and outlier rate parameters on the basis of the original gray wolf algorithm, and uses a new search equation to improve the algorithm's global optimization ability and multi-peak function optimization ability. Since the Ω wolf has the lowest status in the wolf pack and is often the wolf that has been expelled or just accepted by the population, the Ω wolf will exist in three states, namely stable state, leaving state, and joining state. Based on this, a dynamic population strategy is proposed, setting two parameters, the population dynamic rate and the outlier rate. The population dynamic rate controls whether the Ω wolf individual is in a stable state or other states, and the outlier rate controls whether the Ω wolf is in a leaving state or a joining state. In the gray wolf optimization algorithm for dynamic populations,

[0082]

[0083]

[0084]

[0085] in is a coefficient matrix, and its values ​​are shown in equations (32)-(34). is the position vector of α, β, δ wolves, is the position vector of the gray wolf, are the intermediate values ​​of the impact of α, β, and δ wolves on the iteration results of Ω wolf;

[0086]

[0087]

[0088]

[0089] in, is a random number matrix with element values ​​ranging from [0,1]. The effect of the position vectors of A, β, and δ wolves on the rank mechanism of Ω wolf is as follows:

[0090]

[0091]

[0092]

[0093]

[0094]

[0095]

[0096] in is the effect of the position vectors of α, β, and δ wolves on the hierarchical mechanism of Ω wolf, is a random number matrix with element value range [0,1], P1 is set as the population dynamic rate, P2 is set as the outlier rate, then the update formula of Ω wolf in the dynamic population gray wolf optimization algorithm is:

[0097]

[0098] in represents the optimized value, q and ω are random numbers in the range [0,1], and is the upper and lower limit matrix of the optimization space, R1 is the population dynamic threshold, and R2 is the outlier threshold;

[0099] In the boundary optimization algorithm based on the dynamic population gray wolf algorithm, the connection between the dynamic population gray wolf algorithm and sub-salient target detection is established through the cost function. The formula of the cost function is:

[0100]

[0101]

[0102]

[0103]

[0104] C c =‖U color -((f alpha ·F color )+(1-f alpha )·B color )‖2 (46)

[0105]

[0106]

[0107]

[0108] Among them, F color , U color , B color They are the main area, the area to be optimized, and the background area corresponding to the color image. s , U s , B s The main area, the area to be optimized, and the background area after the image optimization mask sampling, F md , B md are the distance matrices between the pixels in the area to be optimized and the pixels in the main area and the background area respectively;

[0109] The dynamic population grey wolf algorithm is used to refine the boundary of the image optimization mask, and a saliency map with uniform target area and clear and complete boundaries is obtained.

[0110] The characteristics and beneficial effects of the present invention are:

[0111] 1) The present invention provides a salient object detection method based on heuristic boundary optimization. Compared with the existing image unsupervised salient object detection algorithm, the image obtained by the present invention has clearer edges and richer details. Compared with the previous methods based solely on extracting manual features or prior knowledge, the method based on heuristic boundary optimization of the present invention has better boundary extraction ability.

[0112] 2) The present invention provides a salient target detection method based on heuristic boundary optimization. Compared with the existing methods, the method of the present invention is an unsupervised method with lower computational complexity, which can effectively reduce the computational cost and data annotation cost.

[0113] 3) The present invention provides a salient target detection method based on heuristic boundary optimization. Compared with the existing methods, this method can effectively improve the integrity of the detected salient target area, and this method has better robustness and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0114] Figure 1 The figure is an overall flow chart of a salient object detection method based on heuristic boundary optimization.

[0115] Figure 2 Schematic diagram for constructing a boundary optimization pyramid using a color image and an image optimization mask. DETAILED DESCRIPTION

[0116] The present invention provides a salient object detection method based on heuristic boundary optimization, the main process of which is as follows: Figure 1 As shown, the main steps are as follows:

[0117] Step 1: Obtain a coarse saliency map based on simple linear iterative clustering and basic digital image features.

[0118] Step 2: Apply the region partitioning algorithm based on inter-class and intra-class variance to obtain the image optimization mask. Perform region partitioning on the coarse saliency map obtained in step 1 to obtain an image optimization mask having three parts: the main region, the region to be optimized, and the background region.

[0119] Step 3: Use the color image and image optimization mask to construct a boundary optimization pyramid. By performing sampling operations with the same parameters on the color image and the image optimization mask at the same time, a multi-scale representation of the image, namely the boundary optimization pyramid, is constructed.

[0120] Step 4: Apply the boundary optimization algorithm based on the dynamic population gray wolf algorithm to optimize the boundary area. The dynamic population-based gray wolf optimization algorithm sets the population dynamic rate and outlier rate parameters on the basis of the original gray wolf algorithm, and uses a new search equation to improve the algorithm's global optimization ability and multi-peak function optimization ability. By refining the boundary of the image optimization mask, a saliency map with uniform target area and clear and complete boundaries is obtained.

[0121] The detailed steps of the present invention are described as follows:

[0122] The present invention mainly solves the problem that the saliency map obtained by the unsupervised saliency target detection method has obvious block phenomenon and poor edge details. In order to solve the above technical problems, the present invention provides a saliency target detection method based on heuristic boundary optimization, and its main steps are as follows:

[0123] Step 1: Obtain a coarse saliency map based on simple linear iterative clustering and basic digital image features. The image is segmented into multiple regions, namely superpixel regions, using the Simple Linear Iterative Cluster (SLIC) method. Within the superpixel region, pixels are grouped or clustered according to human visual perception, and multiple homogeneous regions can be extracted in an image. Specifically, the number of colors is reduced based on color homogeneity combined with color quantization. Afterwards, the feature contrast map is extracted by measuring the color contrast and spatial contrast between superpixel regions. In addition, the input image is subjected to corner detection using the Harris operator to obtain the salient points in the input image. On this basis, the spatial distribution of superpixel regions is used to highlight salient targets and suppress background. Afterwards, the probability of each superpixel region being the background is determined, and the background connectivity is represented by a numerical value. The feature contrast map, spatial compactness term, and background correlation term are smoothed separately and fused together to generate an initial salient feature map. Next, the cost function is used to fuse the initial salient feature map and the background probability map to obtain a coarse saliency map.

[0124] Step 2: Apply a regional partitioning algorithm based on inter-class and intra-class variance to obtain the image optimization mask. Apply a regional partitioning algorithm based on inter-class and intra-class variance to divide the coarse saliency map obtained in step 1 into three parts: the main area, the area to be optimized, and the background area. Use the criteria of maximum inter-class variance and minimum intra-class variance to construct a cost function and obtain the image optimization mask segmentation thresholds T1 and T2. Segment the coarse saliency prediction map based on T1 and T2, remove noise through an open operation, and obtain the image optimization mask.

[0125] Step 3: Use the color image and image optimization mask to construct a boundary optimization pyramid. Perform sampling operations with the same parameters on the color image and the image optimization mask at the same time to construct a multi-scale representation of the image, namely the boundary optimization pyramid. To prevent the boundary optimization pyramid from containing too small images, set the boundary optimization pyramid establishment stop criterion. When the boundary optimization pyramid establishment stop criterion is met, the algorithm stops calculating and obtains the boundary optimization pyramid data model.

[0126] Step 4: Apply the boundary optimization algorithm based on the dynamic population gray wolf algorithm to optimize the boundary area. Before the iterative search, the gray wolf optimization algorithm based on the dynamic population is used for optimization, which provides the initial pixel pairs for heuristic information propagation. For the boundary optimization pyramid data model, starting from the smallest scale image, a multi-scale pixel pair search for heuristic information propagation is performed step by step, and the heuristic information obtained from the previous layer of the pyramid is propagated to the current layer, and heuristic information propagation and single-scale search are performed alternately. Among them, the dynamic population gray wolf optimization algorithm sets the population dynamic rate and outlier rate parameters on the basis of the original gray wolf algorithm, and uses the new search equation to improve the algorithm's global optimization ability and multi-peak function optimization ability. By refining the boundary of the image optimization mask, a saliency map with uniform target area and clear and complete boundaries is obtained.

[0127] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0128] like Figure 1 FIG. 1 is an overall flow chart of a salient object detection method based on heuristic boundary optimization of the present invention. The flow chart specifically includes the following steps:

[0129] Step 1: Obtain a coarse saliency map based on simple linear iterative clustering and basic digital image features. The Simple Linear Iterative Cluster (SLIC) method is used to segment the image into multiple regions, namely superpixel regions. Within the superpixel region, pixels are grouped or clustered according to human visual perception, and multiple homogeneous regions can be extracted in an image. Specifically, the number of colors is reduced based on color homogeneity combined with color quantization. Each color channel of the input image is quantized and uniformly quantized into q binary numbers, where q is set to 12. From the quantized image, a dominant color is selected to cover more than 95% of the pixels in the image. According to the minimum Euclidean distance criterion, colors that cover less than 5% of the pixels are merged into the selected dominant color. After superpixel-based image segmentation and color quantization, each superpixel region S i (i=1,2,…n) all have their histogram H i , and Normalize, where k is the color quantization number. Then, the feature contrast map is extracted by measuring the color contrast and spatial contrast between superpixel regions. i and S j , their color distance C d (S i ,S j ) is defined as the Euclidean distance between the average colors of two superpixel regions, ranging from [0,1]. Here the color distance C dEvaluate the color similarity of two superpixel regions. In addition to the color similarity, i and S j The Gestalt law of proximity is expressed as the spatial proximity distance P d (S I ,S j ), which is defined as the Euclidean distance between the centroids of two superpixels. d (S I ,S j ) is larger, S i and S j The closer. The spatial proximity distance can be further normalized within [0,1] as The formula is:

[0130]

[0131] Where P d-min and P d-max P d The minimum and maximum of all possible values,determined by the superpixels of a given image.

[0132] Based on color distance C d and standardized spatial proximity The global superpixel color contrast of Si is obtained, which is defined as:

[0133]

[0134] Where n is the number of superpixel regions, A j For super pixel S j The normalized area of ​​​​the large S i The global color contrast value of Aj contributes more, and the sum of Aj is 1.

[0135] Similarly, define the superpixel region S i The global spatial contrast values ​​are as follows:

[0136]

[0137] Where L j is the weight of the spatial layout, which is defined as the superpixel region S j The minimum distance from the center of mass of the image to any of the four borders of the image; C I is the color similarity between superpixel regions defined by equation (4), where Hi and Hj refer to two superpixel regions S i and S j The color histogram of , m is the index of the color quantization number k.

[0138]

[0139] In the formula, by (1-|H i (m)-H j (m)|) measures the similarity between two histograms based on their frequencies, which improves the similarity measure and is more consistent with human perception than the traditional histogram intersection method.

[0140] For simplicity, ICS is defined as the superpixel region S i and S j A combined measure of color similarity and spatial proximity between:

[0141] ICS(S i ,S j )=P d-norm (S i ,S j )·C I (S i ,S j ) (5)

[0142] Therefore, the formula for the global spatial contrast value can be simplified to:

[0143]

[0144] In the global color contrast C G and the global spatial contrast S G Based on this, we get the super pixel area S i The characteristic contrast value FC(S i ), whose formula is:

[0145] FC(S i )=C G (S i )·S G (S i ) (7)

[0146] In addition, the input image is corner-detected by Harris operator to obtain the salient points in the input image. On this basis, the spatial distribution of superpixel regions is used to highlight salient targets and suppress background.

[0147] For a given superpixel region image, the colors of all superpixel regions can also be divided into 8 coarse clusters in the Lab color space, where the 3 individual color components are separated by the threshold T (.) For each of the 8 color clusters, the extracted main color is recorded as in is the average value of one of the three color components of Lab, C iAs the center of the corresponding color cluster. The Harris operator is used to determine the coarse area of ​​the foreground and background, and the convex hull is calculated to surround all salient points. All superpixel areas within the convex hull are defined as the foreground, and the remaining superpixels are defined as the background. and Represent the average color of the foreground Gf and background Gb respectively. In order to maximize the color contrast, in each color component, the adaptive threshold T (.) is defined as:

[0148]

[0149] Based on the 8 main colors adaptively determined above, similar colors are merged to simplify subsequent calculations. Compare each pair of extracted main colors, if their Euclidean distance is less than the threshold T d , then merge them. For each image, the threshold is adaptively determined as and half of the Euclidean distance between d The upper limit is set to 15.

[0150] Let Ci and Cj be two color clusters and merge them using weighted average agglomeration method:

[0151]

[0152] Where pi and pj represent C i and C j For each superpixel region image, 3-8 main colors can usually be extracted, which are located in the background or foreground, that is, up to 16 color clusters are obtained, half of which are from the foreground and the others are from the background. As spatial information is not included in color quantization, the superpixels within each color cluster are not necessarily grouped together spatially. This means that each color cluster may contain multiple independent objects consisting of one or more superpixels. Therefore, the superpixel region image becomes an object region image. For G f or G b Each color cluster C in k ,M(C k ) represents its geographical center. k The spatial compactness term is defined as M(C k ) with each object O i The average distance between , which contains:

[0153]

[0154]

[0155] Where k≤8, N k Represents color cluster C kThe number of objects in M(O i ) represents the geographic center of the object O i . Through formula (12), Sd can be scaled to [0,1].

[0156]

[0157] Where Sd(C k ) is Sd f (C k ) or Sd b (C k );Sd min and Sd max are the minimum and maximum values ​​of all spatial compactness Sd, respectively.

[0158] The background correlation term is calculated to measure the degree of connection between object O and the image boundary. A larger value indicates a higher degree of overlap between O and the image boundary. Therefore, since it is assumed that prominent objects should be far away from the image boundary, object O is more likely to be classified as background. The background correlation term is defined as:

[0159]

[0160] Where N c is the total number of superpixels in object O, N b is the number of superpixels in object O that are directly connected to the image boundary. c The maximum and minimum values ​​of all possible values ​​of are denoted by B c_max and B c_min , we can get the normalized B in the range of [0,1] c , the formula is:

[0161]

[0162] After that, the probability of each superpixel region being the background is determined, and the background connectivity is represented by a numerical value. i , the probability that it is the background can be expressed by formula (15):

[0163]

[0164]

[0165]

[0166] Where σ is set to 10, δ(·) is 1 when the superpixel region is on the image boundary, otherwise δ(·) is 0. n is the number of superpixel regions. D geo (S i ,S j) is the geodesic distance between any two superpixel regions, which is determined by the cumulative weight d app It is obtained by following the adjacent superpixel area S in the figure i and S j The shortest path between sequences is measured by the Euclidean distance, where l is the total number of superpixel regions in the determined path.

[0167]

[0168] In order to make BP in [0,1], σ BP Set to 1, the normalization process is defined as:

[0169]

[0170] The feature contrast map, spatial compactness term, and background correlation term are smoothed separately and fused together to generate the initial salient feature map. For each superpixel, its saliency value is replaced by the weighted average of the saliency values ​​of the surrounding superpixel regions. For a given superpixel region S i , select The nearest neighbor is weighted averaged and the local refinement S i The characteristic contrast value is as follows:

[0171]

[0172] in It is in S i The sum of the color similarity and spatial similarity between it and its k nearest neighboring regions.

[0173] For C k All superpixel regions in will be assigned the same spatial compactness term N Sd (C k ). The lower this value is, the higher its similarity to the salient object is. After that, the spatial compactness term of each superpixel region Si is smoothed, and its formula is as follows:

[0174]

[0175] Where n is the number of superpixel regions, YesS i The sum of color similarity and spatial proximity to other superpixel regions.

[0176] In addition, the background related term can also be calculated by formula (22):

[0177]

[0178] Then, the space compactness term R Sd (Si ) and background related items Combined, determine the background suppression map:

[0179]

[0180] The initial saliency feature map can be expressed as:

[0181] ISA(S i )=RFC(S i )·OP(S i ) (twenty four)

[0182] Then, the cost function is used to fuse the initial saliency feature map and the background probability map to obtain a rough saliency map. i is the super pixel area S i The final significance value of . The cost function is

[0183]

[0184]

[0185] where σ is set to 10.

[0186] Step 2: Apply the region partitioning algorithm based on inter-class and intra-class variance to obtain the image optimization mask. Based on the region partitioning based on variance, the coarse saliency map obtained in step 1 is divided into three parts: the main area, the area to be optimized, and the background area. Assume that the image thresholds are T1 and T2, and the global average is m g , pixels less than T1 are background, with an average gray value of u0; pixels greater than T1 and less than T2 are areas to be optimized, with an average gray value of u1; pixels greater than T2 are target areas, with an average gray value of u2. The two thresholds are obtained using the criteria of maximum inter-class variance and minimum intra-class variance. The inter-class variance calculation formula is:

[0187] g inter =ω0×(u0-m g ) 2 +ω1×(u1-m g ) 2 +ω2×(u2-m g ) 2 (26)

[0188] Among them, ω0=3, ω1=1, ω2=3 are the weight coefficients of the three regions respectively. Since it is easier to obtain more accurate reasoning in the main area and the background area, the weight coefficients corresponding to the two areas are larger. g is the global average gray value.

[0189] The formula for calculating the intra-class variance is:

[0190]

[0191] Among them, m0, m1, and m2 are the number of pixels in the main area, the area to be optimized, and the background area respectively. i , x j , x k The pixel values ​​of the three areas are respectively.

[0192] The threshold segmentation cost function is constructed based on the inter-class variance and intra-class variance, and its formula is:

[0193] g=α g ×g inter ×β g -(1-α g )×g intra (28)

[0194] Among them, α g represents the inter-class variance ratio coefficient, α g =0.5,β g represents the inter-class variance weight coefficient, β g =0.14. Adjust the values ​​of T1 and T2. When g reaches the maximum value, T1 and T2 are the optimal thresholds. Segment the coarse saliency prediction map based on T1 and T2, and remove the noise through opening operation to obtain the image optimization mask. In the image optimization mask, the pixel value of the main area is 255, the pixel value of the area to be optimized is 128, and the pixel value of the background area is 0.

[0195] Step 3: Use the color image and image optimization mask to construct a boundary optimization pyramid. Perform sampling operations with the same parameters on the color image and the image optimization mask at the same time to construct a multi-scale representation of the image, namely the boundary optimization pyramid, as shown in Figure 2 As shown in the figure, the images on two adjacent levels of the boundary optimization pyramid are very similar. Considering the similarity, searching for pixel pairs in small-scale images can provide heuristic information for searching large-scale images. To prevent the boundary optimization pyramid from containing too small images, a stopping criterion for establishing the boundary optimization pyramid is set. For example, when one of the length or width of the image drops below 100 pixels, the downsampling loop terminates. The image size set in the stopping criterion is not a sensitive parameter and can be adjusted according to the actual image. When the stopping criterion for establishing the boundary optimization pyramid is met, the algorithm stops calculating and obtains the boundary optimization pyramid data model.

[0196] Step 4: Apply the boundary optimization algorithm based on the dynamic population gray wolf algorithm to optimize the boundary area. Before the iterative search, the dynamic population-based gray wolf optimization algorithm is used to optimize, which provides the initial pixel pairs for heuristic information propagation. For the boundary optimization pyramid data model, starting from the smallest scale image, multi-scale pixel pair search for heuristic information propagation is performed step by step, the heuristic information obtained from the previous layer of the pyramid is propagated to the current layer, and the heuristic information propagation and single-scale search are performed alternately.

[0197] In heuristic information propagation, a boundary optimization pyramid is used to define the neighborhood of pixel pairs on adjacent levels. The pixels in the current layer of the pyramid, which are used to calculate the average color of the pixels in the previous layer, are defined as the neighborhood of the pixels in the previous layer. The neighborhood relationship of pixels of different scales can be extended to pixel pairs. Some pixel pairs in the initial decision vector obtained by heuristic information propagation have higher quality due to their local smoothness prior. In the pixel pair selection process, the optimization algorithm is not applied to high-quality pixel pairs, and only half of the low-quality pixel pairs are selected from the initial decision vector to be optimized to avoid redundant search.

[0198] The gray wolf optimization algorithm for dynamic populations sets the population dynamic rate and outlier rate parameters based on the original gray wolf algorithm, and uses a new search equation to enhance the algorithm's global optimization capability and multi-peak function optimization capability. Since the Ω wolf has the lowest status in the wolf pack, it is often the wolf that has been expelled or just accepted by the pack. Therefore, the Ω wolf will exist in three states, namely the stable state, the leaving state, and the joining state. Based on this, a dynamic population strategy is proposed, setting two parameters: the population dynamic rate and the outlier rate. The population dynamic rate controls whether the Ω wolf individual is in a stable state or other states, and the outlier rate controls whether the Ω wolf is in a leaving state or a joining state. In the gray wolf optimization algorithm for dynamic populations,

[0199]

[0200]

[0201]

[0202] in is a coefficient matrix, and its values ​​are shown in equations (32)-(34). is the position vector of α, β, δ wolves, is the position vector of the gray wolf, They are the intermediate values ​​of the impact of α, β, and δ wolves on the iterative results of Ω wolf.

[0203]

[0204]

[0205]

[0206] in, is a random number matrix whose elements range from [0,1]. The effect of the position vectors of A, β, and δ wolves on the rank mechanism of Ω wolf is as follows:

[0207]

[0208]

[0209]

[0210]

[0211]

[0212]

[0213] in is the effect of the position vectors of α, β, and δ wolves on the hierarchical mechanism of Ω wolf, is a random number matrix whose elements range from [0,1]. Assume P1 is the population dynamic rate and P2 is the outlier rate. Then the update formula of Ω wolf in the gray wolf optimization algorithm of dynamic population is:

[0214]

[0215] in represents the optimized value, q and ω are random numbers in the range [0,1], and To optimize the upper and lower limit matrices of the space, R1 is the population dynamics threshold, and R2 is the outlier threshold.

[0216] In the boundary optimization algorithm based on the dynamic population gray wolf algorithm, the connection between the dynamic population gray wolf algorithm and sub-salient target detection is established through the cost function. The formula of the cost function is:

[0217]

[0218]

[0219]

[0220]

[0221] C c =‖U color -((f alpha ·F color )+(1-f alpha )·Bcolor )‖2 (46)

[0222]

[0223]

[0224]

[0225] Among them, F color , U color , B color They are the main area, the area to be optimized, and the background area corresponding to the color image. s , U s , B s The main area, the area to be optimized, and the background area after the image optimization mask sampling, F md , B md They are the distance matrices between the pixels in the area to be optimized and the pixels in the main area and the background area.

[0226] The dynamic population grey wolf algorithm is used to refine the boundary of the image optimization mask, and a saliency map with uniform target area and clear and complete boundaries is obtained.

[0227] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed in the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for detecting salient objects based on heuristic boundary optimization, characterized in that the steps as follows: Step 1: Obtain a coarse saliency map based on simple linear iterative clustering and basic digital image features; Step 2: Apply the region partitioning algorithm based on inter-class and intra-class variance to obtain the image optimization mask. Perform region partition on the coarse saliency map obtained in step 1 to obtain an image optimization mask with three parts: the main area, the area to be optimized, and the background area. Assume that the image thresholds are T1 and T2, and the global average is m g , pixels less than T1 are background, with an average gray value of u0; pixels greater than T1 and less than T2 are areas to be optimized, with an average gray value of u1; pixels greater than T2 are target areas, with an average gray value of u2. The two thresholds are obtained using the criteria of maximum inter-class variance and minimum intra-class variance. The inter-class variance calculation formula is: g inter =ω0×(u0-m g ) 2 +ω1×(u1-m g ) 2 +v2×(u2-m g ) 2 (26) Among them, ω0=3, ω1=1, ω2=3 are the weight coefficients of the three regions respectively. Since it is easier to obtain more accurate reasoning in the main area and the background area, the weight coefficients corresponding to the two areas are larger, m g is the global average gray value; the intra-class variance calculation formula is: Among them, m0, m1, and m2 are the number of pixels in the main area, the area to be optimized, and the background area respectively, and x i , x j , x k Then the pixel values ​​of the three regions are respectively; The threshold segmentation cost function is constructed based on the inter-class variance and intra-class variance, and its formula is: g=a g ×g inter ×b g -(1-a g )×g intra (28) Among them, α g represents the inter-class variance ratio coefficient, α g =0.5,β g Represents the inter-class variance weight coefficient, adjusts the values ​​of T1 and T2, when g reaches the maximum value, T1 and T2 are the optimal thresholds, the coarse saliency prediction map is segmented based on T1 and T2, and the noise is removed through the opening operation to obtain the image optimization mask; Step 3: Use the color image and the image optimization mask to construct a boundary optimization pyramid. By performing sampling operations with the same parameters on the color image and the image optimization mask at the same time, a multi-scale representation of the image, namely, a boundary optimization pyramid, is constructed. Step 4: Apply the boundary optimization algorithm based on the dynamic population gray wolf algorithm to optimize the boundary area. The dynamic population-based gray wolf optimization algorithm sets the population dynamic rate and outlier rate parameters on the basis of the original gray wolf algorithm, and uses the new search equation to improve the global optimization ability and multi-peak function optimization ability of the algorithm. By refining the boundary of the image optimization mask, a saliency map with uniform target area and clear and complete boundaries is obtained.

2. The method for detecting salient objects based on heuristic boundary optimization as claimed in claim 1, characterized in that: The specific steps are as follows: Step 1: A coarse saliency map is obtained based on simple linear iterative clustering and basic digital image features. The simple linear iterative clustering (SLIC) method is used to segment the image into multiple regions, namely superpixel regions. Within the superpixel region, pixels are grouped or clustered according to human visual perception. Multiple homogeneous regions can be extracted from an image. After that, a feature contrast map is extracted by measuring the color contrast and spatial contrast between superpixel regions. In addition, the input image is subjected to corner detection by the Harris operator to obtain the salient points in the input image. On this basis, the spatial distribution of superpixel regions is used to highlight the salient targets and suppress the background. After that, the probability of each superpixel region being the background is determined, and the background connectivity is represented by a numerical value. The feature contrast map, the spatial compactness term, and the background correlation term are smoothed separately and fused together to generate an initial salient feature map. Then, the cost function is used to fuse the initial salient feature map and the background probability map to obtain a coarse saliency map. Step 2: Apply the regional division algorithm based on inter-class and intra-class variance to obtain the image optimization mask: Apply the regional division algorithm based on inter-class and intra-class variance to divide the coarse saliency map obtained in step 1 into three parts: the main area, the area to be optimized, and the background area. Use the criteria of maximum inter-class variance and minimum intra-class variance to construct the cost function, calculate the image optimization mask segmentation thresholds T1 and T2, segment the coarse saliency prediction map based on T1 and T2, and remove noise through opening operation to obtain the image optimization mask; Step 3: Use the color image and the image optimization mask to construct a boundary optimization pyramid. At the same time, perform sampling operations with the same parameters on the color image and the image optimization mask to construct a multi-scale representation of the image, namely, the boundary optimization pyramid. To avoid the boundary optimization pyramid from containing too small an image, a boundary optimization pyramid establishment stop criterion is set. When the boundary optimization pyramid establishment stop criterion is met, the algorithm stops calculating and obtains a boundary optimization pyramid data model. Step 4: Apply the boundary optimization algorithm based on the dynamic population gray wolf algorithm to optimize the boundary area. Before the iterative search, the gray wolf optimization algorithm based on the dynamic population is used for optimization. The algorithm provides the initial pixel pairs for heuristic information propagation. For the boundary optimization pyramid data model, starting from the smallest scale image, multi-scale pixel pair search for heuristic information propagation is performed step by step, and the heuristic information obtained from the previous layer of the pyramid is propagated to the current layer, and heuristic information propagation and single-scale search are performed alternately; Among them, the dynamic population gray wolf optimization algorithm sets the population dynamic rate and outlier rate parameters on the basis of the original gray wolf algorithm, and uses the new search equation to improve the algorithm's global optimization ability and multi-peak function optimization ability. By refining the boundary of the image optimization mask, a saliency map with uniform target area and clear and complete boundaries is obtained.

3. The method for detecting salient objects based on heuristic boundary optimization as claimed in claim 1, characterized in that: The detailed steps are as follows: Step 1: Obtain a coarse saliency map based on simple linear iterative clustering and basic digital image features: Use the simple linear iterative clustering SLIC (Simple Linear Iterative Cluster) method to segment the image into multiple regions, namely superpixel regions. Within the superpixel region, pixels are grouped or clustered according to human visual perception; quantize each color channel of the input image and quantize it uniformly into q binary numbers. From the quantized image, select the main color to cover more than 95% of the pixels in the image. According to the minimum Euclidean distance criterion, the color covering less than 5% of the pixels is merged into the selected main color. After image segmentation and color quantization based on superpixels, each superpixel region S i Each has its histogram H i , i = 1, 2, ... n, and Normalize, where k is the color quantization number; then, extract the feature contrast map by measuring the color contrast and spatial contrast between superpixel regions. i and S j , their color distance C d (S i ,S j ) is defined as the Euclidean distance between the average colors of two superpixel regions, ranging from [0,1]. Here, the color distance C d Evaluate the color similarity of two superpixel regions. In addition to the color-based similarity, we also evaluate the color similarity of two superpixels S. i and S j The Gestalt law of proximity is expressed as the spatial proximity distance P d (S I ,S j ), which is defined as the Euclidean distance between the centroids of two superpixels, P d (S I ,S j ) is larger, S i and S j The closer the distance is, the closer the spatial distance can be further normalized in [0,1] as The formula is: Where P d_min and P d_max P d The minimum and maximum of all possible values, determined by the superpixels of a given image; Based on color distance C d and standardized spatial proximity The global superpixel color contrast of Si is obtained, which is defined as: Where n is the number of superpixel regions, A j For super pixel S j The normalized area of ​​​​the large S i The global color contrast value of contributes more, and the sum of Aj is 1; Define the superpixel region S i The global spatial contrast values ​​are as follows: Where L j is the weight of the spatial layout, which is defined as the superpixel region S j The minimum distance from the center of mass of the image to any of the four borders of the image; C I is the color similarity between superpixel regions defined by equation (4), where Hi and Hj refer to two superpixel regions S i and S j The color histogram of , m is the index of the color quantization number k; In the formula, by (1-|H i (m)-H j (m)|) measures the similarity of two histograms based on frequency; ICS is defined as the superpixel region S i and S j A combined measure of color similarity and spatial proximity between: ICS(S i ,S j )=P d_norm (S i ,S j )·C I (S i ,S j ) (5) Therefore, the formula for the global spatial contrast value can be simplified to: In the global color contrast C G and the global spatial contrast S G Based on this, we get the super pixel area S i The characteristic contrast value FC(S i ), whose formula is: FC(S i )=C G (With i )·S G (With i ) (7) In addition, the input image is subjected to corner detection through the Harris operator to obtain the salient points in the input image. On this basis, the spatial distribution of the superpixel area is used to highlight the salient target and suppress the background. For a given superpixel region image, the colors of all superpixel regions can also be divided into 8 coarse clusters in the Lab color space, where the 3 individual color components are separated by the threshold T (.) Divided into two parts, for each of the 8 color clusters, the extracted main color is recorded as in is the average value of one of the three color components of Lab, C i As the center of the corresponding color cluster, the Harris operator is used to determine the coarse area of ​​the foreground and background, and the convex hull is calculated to surround all salient points. All superpixel areas within the convex hull are defined as the foreground, and the remaining superpixels are defined as the background. and Represent the average color of the foreground Gf and background Gb respectively. In order to maximize the color contrast, in each color component, the adaptive threshold T (.) is defined as: According to the 8 main colors adaptively determined above, similar colors are merged to simplify subsequent calculations. Each pair of extracted main colors is compared. If their Euclidean distance is less than the threshold T d , then merge them, and for each image, adaptively determine the threshold as and Half of the Euclidean distance between them, let Ci and Cj be two color clusters, and merge them using the weighted average agglomeration method: Where pi and pj represent C i and C j The number of pixels in G f or G b Each color cluster C in k ,M(C k ) represents its geographical center, C k The spatial compactness term is defined as M(C k ) with each object O i The average distance between , which contains: Where k≤8, N k Represents color cluster C k The number of objects in M(O i ) represents the geographic center of the object O i , through formula (12), Sd is scaled to [0,1]; Where Sd(C k ) is Sd f (C k ) or Sd b (C k );Sd min and Sd max are the minimum and maximum values ​​of all spatial compactness Sd, respectively; The background correlation term is calculated to measure the degree of connection between object O and the image boundary. The larger the value, the higher the overlap between O and the image boundary. The background correlation term is defined as: Where N c is the total number of superpixels in object O, N b is the number of superpixels in object O that are directly connected to the image boundary, and B c The maximum and minimum values ​​of all possible values ​​of are denoted by B c_max and B c_min , get the normalized B in the range [0,1] c , the formula is: After that, the probability of each superpixel region being the background is determined, and the background connectivity is represented by a numerical value. For a given superpixel region S i , the probability that it is the background can be expressed by formula (15): Where σ is set to 10, δ(·) is 1 when the superpixel region is on the image boundary, otherwise δ(·) is 0, n is the number of superpixel regions, and D geo (S i ,S j ) is the geodesic distance between any two superpixel regions, which is determined by the cumulative weight D app It is obtained by following the adjacent superpixel area S in the figure i and S j The Euclidean distance of the shortest path between sequences is measured, and l is the total number of superpixel regions in the determined path; In order to make BP in [0,1], σ BP Set to 1, the normalization process is defined as: The feature contrast map, spatial compactness term, and background correlation term are smoothed separately and fused together to generate the initial salient feature map. For each superpixel, its saliency value is replaced by the weighted average of the saliency values ​​of the surrounding superpixel regions. For a given superpixel region S i , select The nearest neighbor is weighted averaged and the local refinement S i The characteristic contrast value is as follows: in It is in S i The sum of the color similarity and spatial similarity between it and its k nearest neighboring regions; For C k All superpixel regions in will be assigned the same spatial compactness term N Sd (C k ), the lower this value is, the higher its similarity with the salient object is. After that, the spatial compactness term of each superpixel region Si is smoothed, and its formula is as follows: Where n is the number of superpixel regions, YesS i The sum of color similarity and spatial proximity to other superpixel regions; In addition, the background related term is calculated by formula (22): Then, the space compactness term R Sd (S i ) and background related items Combined, determine the background suppression map: The initial saliency feature map is expressed as: ISA(S i )=RFC(S i )·OP(S i ) (24) Next, the cost function is used to fuse the initial saliency feature map and the background probability map to obtain a rough saliency map. Let FSA i is the super pixel area S i The final significance value of Step 3: Use the color image and the image optimization mask to construct a boundary optimization pyramid. At the same time, perform sampling operations with the same parameters on the color image and the image optimization mask to construct a multi-scale representation of the image, namely, the boundary optimization pyramid. To avoid the boundary optimization pyramid from containing too small images, set a boundary optimization pyramid establishment stop criterion: when one of the length or width of the image drops below 100 pixels, the downsampling loop terminates. When the boundary optimization pyramid establishment stop criterion is met, the algorithm stops calculating and obtains the boundary optimization pyramid data model. Step 4: Apply the boundary optimization algorithm based on the dynamic population gray wolf algorithm to optimize the boundary area. Before the iterative search, the gray wolf optimization algorithm based on the dynamic population is used for optimization. The algorithm provides the initial pixel pairs for heuristic information propagation. For the boundary optimization pyramid data model, starting from the smallest scale image, multi-scale pixel pair search for heuristic information propagation is performed step by step. The heuristic information obtained from the previous layer of the pyramid is propagated to the current layer, and the heuristic information propagation and single-scale search are performed alternately. In the heuristic information propagation, the boundary optimization pyramid is used to define the neighborhood of pixel pairs on adjacent levels. The pixels in the current layer of the pyramid used to calculate the average color of the pixels in the previous layer are defined as the neighborhood of the pixels in the previous layer. The neighborhood relationship of pixels of different scales can be extended to pixel pairs. In the process of selecting pixel pairs, the optimization algorithm is not applied to high-quality pixel pairs. Only half of the low-quality pixel pairs are selected from the initial decision vector to be optimized to avoid redundant search. The gray wolf optimization algorithm for dynamic populations sets the population dynamic rate and outlier rate parameters on the basis of the original gray wolf algorithm, and uses the new search equation to improve the algorithm's global optimization ability and multi-peak function optimization ability. Since the Ω wolf has the lowest status in the wolf pack and is often the wolf that has been expelled or just accepted by the population, the Ω wolf will exist in three states, namely, stable state, leaving state, and joining state. Based on this, a dynamic population strategy is proposed, and two parameters, population dynamic rate and outlier rate, are set. The population dynamic rate controls whether the Ω wolf individual is in a stable state or other states, and the outlier rate controls whether the Ω wolf is in a leaving state or a joining state. In the gray wolf optimization algorithm for dynamic populations, in is a coefficient matrix, and its values ​​are shown in equations (32)-(34). is the position vector of α, β, δ wolves, is the position vector of the gray wolf, are the intermediate values ​​of the impact of α, β, and δ wolves on the iteration results of Ω wolf; in, is a random number matrix with element values ​​ranging from [0,1]. The effect of the position vectors of α, β, and δ wolves on the level mechanism of Ω wolf is as follows: in is the effect of the position vectors of α, β, and δ wolves on the hierarchical mechanism of Ω wolf, is a random number matrix with element value range [0,1], P1 is set as the population dynamic rate, P2 is set as the outlier rate, then the update formula of Ω wolf in the dynamic population gray wolf optimization algorithm is: in represents the optimized value, q and ω are random numbers in the range [0,1], and is the upper and lower limit matrix of the optimization space, R1 is the population dynamic threshold, and R2 is the outlier threshold; In the boundary optimization algorithm based on the dynamic population gray wolf algorithm, the connection between the dynamic population gray wolf algorithm and sub-salient target detection is established through the cost function. The formula of the cost function is: C c =‖U color -((f alpha ·F color )+(1-f alpha )·B color )‖2 (46) Among them, F color , U color , B color They are the main area, the area to be optimized, and the background area corresponding to the color image. s , U s , B s The main area, the area to be optimized, and the background area after the image optimization mask sampling, F md , B md are the distance matrices between the pixels in the area to be optimized and the pixels in the main area and the background area respectively; The dynamic population grey wolf algorithm is used to refine the boundary of the image optimization mask, and a saliency map with uniform target area and clear and complete boundaries is obtained.

Images