Saliency target detection method based on superpixel segmentation

Abstract

The invention relates to a saliency target detection method based on superpixel segmentation. The method comprises the following steps: obtaining an image to be detected; based on the to-be-detected image, performing super-pixel segmentation by using a super-pixel segmentation algorithm to obtain a segmented image; based on the segmented image, a color enhancement detection algorithm is adopted for processing, and an initial saliency map containing the interest points is obtained; based on the initial saliency map containing the interest points, a convex hull model is adopted for processing, and an initial saliency map containing candidate target detection areas is obtained; and adopting a multi-layer cellular automaton to discriminate the foreground and the background in the initial saliency map containing the candidate target detection area, generating a final saliency map, and obtaining a saliency target. Compared with the prior art, the method has the advantages that a complete salient target area can be obtained, boundary information is refined, and a better detection effect can be obtained.

Description

Technical Field

[0001] The present invention relates to the technical field of target detection, and in particular to a salient target detection method based on superpixel segmentation. Background Art

[0002] Advanced science and technology are gradually being integrated into people's daily lives, constantly impacting their basic needs for food, clothing, housing, and transportation. This has led to an exponential growth in the amount of digital information, from text input to voice input, and from image display to video playback. Computer image processing technology, which corresponds to this data growth, is now urgently needed. Existing computer image processing technology boasts powerful computing capabilities, such as high speed, high precision, and robust logic, capable of processing large amounts of data. However, in this era of data explosion, current computer image processing speeds are far from meeting current needs.

[0003] With the rapid development of internet technology, saliency detection has achieved increasing success in related fields such as image processing and has been widely applied in areas such as image compression, object recognition and tracking, scene classification, and image segmentation. However, most current saliency detection methods often rely on manually defined low-level features and only target single objects with simple background information. This, to a certain extent, limits their effectiveness in complex scenes with multiple objects. Research has shown that the human visual system is able to quickly and accurately locate the most important salient target areas in complex natural images and then respond accordingly. Visual attention plays a crucial role in this process. Due to our limited perceptual and cognitive resources, we are forced to focus our attention on the most useful and interesting parts of these resources. Researchers have proposed an "attention mechanism" based on human vision. This mechanism, called the Human Visual Attention Mechanism (HVS), can quickly and accurately detect regions of interest. Computer vision aims to mimic the human visual system's ability to process natural images, using computers to simulate human vision's ability to automatically extract the most useful information from natural images.

[0004] Saliency detection generally refers to identifying the most interesting regions in an image. Saliency maps are the most intuitive form of image saliency detection. Saliency maps effectively suppress redundant and complex background areas in an image, retaining only the salient regions as the output grayscale image. The task of saliency detection is to quickly extract the target region of interest in an image. The saliency map identifies the flower, which is the region of interest in the entire image, known as the salient target region, while automatically ignoring the complex background areas.

[0005] Current technology mainly focuses on detecting larger targets in images, but the detection of small targets is easily overlooked. When the foreground salient area or the target and background features are similar, the saliency detection effect is poor. However, at present, saliency detection of natural images with similar foreground and background colors has always been a difficult problem. Most existing methods perform saliency detection based on underlying or shallow features such as color and texture. The selection and extraction of these underlying features are largely done manually and are inevitably affected by subjective factors. Depending on the selection of underlying features, different methods often only work well for a certain type of specific image, and are not suitable for multi-target saliency detection in complex scenes. There are currently several designs for salient target detection:

[0006] 1) A saliency detection based on the local contrast method. Ma et al. proposed a new framework for image attention mechanism analysis, which consists of attention perspective, gaze area and gaze point. The model framework can quickly perform saliency detection. Achanta et al. used the inner and outer sub-regions of a sliding rectangular window centered on a pixel to calculate the average feature vector between the two windows and obtain the saliency map of the algorithm. Rahtu et al. proposed a novel method for saliency detection, which is to calculate the saliency value of the algorithm by comparing the central and surrounding features, and the algorithm incorporates the Bayesian formula, in which the conditional probability is represented by the distribution characteristics of color.

[0007] 2) A saliency detection method based on global contrast. Achanta et al. proposed a salient region detection algorithm by preserving frequency content, commonly known as the FT model. The FT model is computationally simple and efficient. This method only uses the color feature value of the pixel to compare with the average color feature value of the entire image, but it is not effective for images with complex backgrounds. Cheng et al. proposed a global contrast salient region detection algorithm using color feature histogram and spatial feature information. Perazzi et al. proposed a salient region detection based on contrast filter (SF for short) considering the characteristics of color and spatial distribution information. The SF model cleverly combines color uniqueness and spatial information characteristics, and uses Gaussian filters to accelerate its calculation process, enabling it to quickly and accurately find the target area. Zhai et al. proposed a spatiotemporal video saliency detection technology. The color histogram of the image is used in the spatial attention model to calculate the saliency map. The saliency value of a pixel is obtained by calculating the color feature distance between a pixel and all other pixels.

[0008] 3) Yan et al. proposed a multi-layered saliency detection method (HS). This method analyzes the target region from a scale perspective. It first extracts three image layers from the input image, then calculates the saliency value of each layer. Finally, the saliency value is input into a hierarchical model to obtain the final saliency map. This method is limited to processing small-scale, high-contrast patterns; in this case, the resulting saliency map is relatively good.

[0009] Saliency detection has achieved a lot of results in the field of computer vision, but it still cannot meet the needs of practical applications. In particular, the technology for collaborative saliency detection of multiple images is not mature enough and there are many problems that need to be solved. By summarizing and analyzing the current research status of saliency detection, it is found that the main problems in the field of saliency detection are as follows: (1) Feature selection and extraction problems. At present, whether it is saliency detection of a single image or collaborative saliency detection of multiple images, they are mostly based on underlying features such as color, space and texture, and are often only effective for a certain type of specific image. The extraction and selection of these underlying features are all done manually, which is a task that requires certain professional knowledge and is extremely labor-intensive. The quality of the selection results mainly depends on subjective experience. Although deep feature extraction methods can be automatically obtained, they are relatively slow, which affects the efficiency of saliency detection. (2) Saliency calculation problems. Most of the existing saliency detection algorithms calculate saliency values based on contrast features. Summary of the Invention

[0010] The purpose of the present invention is to provide a salient object detection method based on superpixel segmentation to improve the accuracy of salient object detection.

[0011] The purpose of the present invention can be achieved by the following technical solutions:

[0012] A salient object detection method based on superpixel segmentation includes the following steps:

[0013] Acquire the image to be tested;

[0014] Based on the image to be tested, superpixel segmentation is performed using a superpixel segmentation algorithm to obtain a segmented image;

[0015] Based on the segmented image, a color enhancement detection algorithm is used to process the image to obtain an initial saliency map containing interest points;

[0016] Based on the initial saliency map containing the interest points, a convex hull model is used to process the initial saliency map to obtain an initial saliency map containing the candidate target detection area;

[0017] A multi-layer cellular automation is used to discriminate the foreground and background in the initial saliency map containing the candidate target detection area, generate a final saliency map, and obtain a salient target.

[0018] Furthermore, the step of obtaining the segmented image includes:

[0019] 1) Initialization parameters, including the number of superpixels K, the number of pixels N and the cluster center C k , where k is the number of cluster centers;

[0020] 2) With the currently set cluster center C k The first neighborhood is set as the center, and the spatial position with the minimum gradient amplitude value is found in the first neighborhood by calculating the image gradient amplitude value as the new cluster center;

[0021] 3) Based on the new cluster center, a second neighborhood is set, and all pixels in the second neighborhood are classified using a nearest neighbor classification method to obtain a classification result;

[0022] 4) Repeat steps 2)-3) to iteratively update the cluster centers and classification results until the iteration ends to obtain a final classification result, which is used as the segmented image.

[0023] Furthermore, the calculation expression of the image gradient amplitude value is:

[0024] G(x,y)=||I(x+1,y)-I(x-1,y)|| 2 +||I(x,y+1)-I(x,y-1)|| 2

[0025] Where G(x,y) is the image gradient amplitude, and I(x,y) is the color vector of the color feature of the pixel at the (x,y) position.

[0026] Furthermore, the step of classifying using the nearest neighbor classification method includes:

[0027] In the second neighborhood, for each pixel, find the new cluster center with the closest distance, and use the category of the new cluster center as the pixel label to complete the pixel classification process. The distance calculation expression is:

[0028]

[0029] Where, d lab and d xy Represents the color feature distance and spatial position feature distance respectively, D s It is d lab and d xyIt is obtained by linear superposition, which represents the feature distance. m represents a variable that controls the compactness between superpixels. l, a, and b represent different channels in the color feature space, and (x, y) represent pixel positions.

[0030] Furthermore, the step of obtaining an initial saliency map containing interest points includes:

[0031] Performing color space conversion based on the segmented image to obtain a converted segmented image;

[0032] Based on the converted segmented image, color saliency enhancement is performed through the conversion function g to generate a saliency matrix M′, wherein the expression of the saliency matrix M′ is:

[0033]

[0034] Among them, the conversion function g satisfies:

[0035]

[0036] I(v)=-log(p(v))

[0037] Where ω(x,y) is the Gaussian weight function, g(I x ) and g(I y ) is the gradient component after color enhancement, P(v) is the probability of event v occurring;

[0038] Based on the saliency matrix M′, local maxima are calculated, all interest points are screened out, and interest points located at the edge are eliminated to obtain an initial saliency map containing interest points.

[0039] Furthermore, the step of obtaining an initial saliency map containing the candidate target detection area includes:

[0040] Based on the initial saliency map containing the interest points, selecting the starting position of the interest points as the starting point of the convex hull, wherein the starting point is located at the lower rightmost position of all the interest points;

[0041] Taking the starting point of the convex hull as a reference, select the interest points with the smallest angle value according to the preset direction and connect them. Then repeat this step with the selected interest point as the next starting point to form a convex hull area as the candidate target detection area, and obtain an initial saliency map containing the candidate target detection area.

[0042] Furthermore, the step of generating the final saliency map includes:

[0043] Each pixel in the initial saliency map containing the candidate target detection area is taken as a cell i, and the initial saliency map containing the candidate target detection area is binarized to obtain a foreground feature saliency map and a background feature saliency map;

[0044] Calculating the prior probability ratio and the posterior probability ratio of each cell i on the foreground feature saliency map and the background feature saliency map;

[0045] Based on the prior probability ratio and the posterior probability ratio, a synchronous update mechanism is used to synchronously update each cell i to adjust the significance value, and this step is repeated for iterative update to obtain the final updated significance value;

[0046] Based on the final updated saliency values of the foreground feature saliency map and the background feature saliency map, a calculation is performed to obtain a final saliency map, wherein the final saliency map is expressed as:

[0047]

[0048] Where, is the final saliency map, N2 is the number of iterative updates, and m is the mth saliency map.

[0049] Furthermore, the step of obtaining a foreground feature saliency map and a background feature saliency map includes:

[0050] Calculate the saliency value S of each cell i in the initial saliency map containing the candidate target detection area i ;

[0051] Set the binarization threshold γ m ;

[0052] Determine each cell i The significance value S i Is it greater than or equal to γ? m If yes, then cell i is marked as foreground, if no, then cell i is marked as background, and finally a foreground feature saliency map and a background feature saliency map are formed;

[0053] Among them, if cell i is marked as foreground, then the probability λ of its neighbor j at the same position on the remaining initial saliency map containing the candidate target detection area being marked as foreground satisfies:

[0054] λ=P(η i =+1|i∈F)

[0055] Where P is the probability function, η i represents the number of neighbor pixels of i that are marked as foreground, F is the number of neighbor pixels i marked as foreground;

[0056] If cell i is marked as background, then the probability μ of neighbor j at the same position on the remaining initial saliency map containing the candidate target detection area being marked as background satisfies:

[0057] μ=P(η i = +1|i∈B);

[0058] Where η i represents the number of neighbor pixels i marked as background, and B represents the number of neighbor pixels i marked as background.

[0059] Furthermore, the expressions of the prior probability ratio and the posterior probability ratio are respectively:

[0060]

[0061] Among them, the l=ln(Λ) transformation is performed on equations (1) and (2) to obtain the final expressions of the prior probability ratio and the posterior probability ratio:

[0062]

[0063] Where Λ(i∈F) is the prior probability ratio, Λ(i∈F|η i =+1) is the posterior probability ratio, λ is the probability that if cell i is marked as foreground, then its neighbor j at the same position on the remaining initial saliency map containing the candidate target detection area is marked as foreground, and μ is the probability that if cell i is marked as background, then its neighbor j at the same position on the remaining initial saliency map containing the candidate target detection area is marked as background.

[0064] Furthermore, the synchronous update mechanism is expressed as:

[0065]

[0066] Where, for, Represents the significance values of all cells on the mth saliency map at time t. Matrix 1 is a matrix with N elements [1,…,1] T , is the significance value of all cells on the kth saliency map at time t, γ k is the binarization threshold of the kth image, λ is the probability that cell i is marked as foreground, and its neighbor j at the same position on the remaining initial saliency map containing the candidate target detection area is marked as foreground.

[0067] Compared with the prior art, the present invention has the following beneficial effects:

[0068] (1) The present invention uses a superpixel segmentation algorithm to perform superpixel segmentation on the image, and then uses a color enhancement detection algorithm to perform color saliency enhancement to find points of interest and combine the convex hull model to delineate the initial saliency map containing the candidate target detection area, and combines the multi-layer cellular automaton to further distinguish the background and foreground. The present invention can obtain a complete salient target area and improve the accuracy of salient target detection.

[0069] (2) In the superpixel segmentation algorithm of the present invention, the cluster center is updated by finding the minimum image gradient amplitude value in the neighborhood, and pixel classification is performed based on the cluster center. The algorithm is repeated multiple times to obtain a better superpixel segmentation effect.

[0070] (3) The present invention adopts a synchronous update mechanism in the multi-layer cellular automation scheme to adjust the saliency value, which can more accurately distinguish the foreground and background in the saliency map, thereby further improving the detection accuracy of salient targets.

[0071] (4) During the process of searching for points of interest, the present invention performs spatial transformation to eliminate the influence of external factors, and uses a saliency matrix to screen points of interest, thereby improving the accuracy of point of interest detection, thereby obtaining a more accurate detection area of the candidate target. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 Schematic diagram of the method flow of the present invention;

[0073] Figure 2 This is the experimental result of the present invention. DETAILED DESCRIPTION

[0074] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0075] This embodiment provides a salient object detection method based on superpixel segmentation. Figure 1 As shown, the method includes the following steps:

[0076] Step 1: Obtain the image to be tested and perform superpixel segmentation.

[0077] For the acquired image, the first step is to set the number of superpixels to K, the number of pixels to N, and the cluster center to C. k , so we can see that the distance between cluster centers is The second step is to reselect the location of the cluster center: by setting C k The location is the center, and C is reselected in the 3*3 neighborhood of its center. kPosition. Find the spatial position with the minimum image gradient amplitude value in this neighborhood as the location of the new cluster center. The purpose is to ensure the accuracy of the selected sampling points and avoid inaccurate results. The image gradient amplitude calculation formula is as follows:

[0078] G(x,y)=||I(x+1,y)-I(x-1,y)|| 2 +||I(x,y+1)-I(x,y-1)|| 2

[0079] Where I(x,y) represents the color vector of the color feature of the pixel at the (x,y) position.

[0080] The third step is iterative optimization: repeat the following operation process until convergence. Classify all pixels in the 2s*2s area around the cluster center, use the nearest neighbor classification method to find the cluster center closest to this pixel, use the category of the cluster center as the label value of this pixel, and end the clustering process. Get the new cluster center from the second step, and re-cluster with the new cluster center. It can be seen that the cluster center needs to be continuously updated and the clustering process is repeated. The iteration is completed until the distance between two adjacent cluster centers is less than the constant value set by the algorithm. Generally speaking, only 10 iterations are required to achieve good superpixel segmentation results. The cluster center is C k , D s is the color feature distance d lab and spatial position characteristic distance d xy It is obtained by linear superposition. Set the cluster center of K superpixels as C k =[l k a k b k x k y k ] T Where K ranges from 1 to K, L, a, and b range from fixed values, and the distance calculation formula is as follows:

[0081]

[0082] Where m represents a variable that controls the compactness between superpixels. lab and d xy Represents the color feature distance and spatial position feature distance respectively, D s Represents the feature distance, and its purpose is to filter which pixels in the area belong to the category of the cluster center.

[0083] The fourth step is to enhance the connectivity of the clustering results: To solve the problems that may arise in the third step, such as a few superpixels are too small, a small number of pixels with the same class label are not in a connected domain, etc., the superpixel labels of more pixels in the disconnected small-size image are re-labeled.

[0084] Step 2: Find the points of interest and obtain the detection area of the candidate target.

[0085] By using the interest point detection algorithm to find all the interest points, we can find that the distribution of all interest points is characterized by the majority being located around the salient target area. However, we cannot rule out the possibility that the interest points are located at the edge of the image. In this case, we need to remove the interest points at the edge of the image to obtain more accurate interest points. First, the input color image is converted to a color space to eliminate the influence of external factors. If the probability of event v occurring is P(v), then the information content of event v is expressed as:

[0086] I(v)=-log(p(v))

[0087] The detection algorithm based on color enhancement proposes to calculate corner points through a conversion function, which requires a conversion function g;

[0088]

[0089] The transformation achieved by the function g is called color saliency enhancement. Once the function g is found, the saliency of the color enhancement can be calculated. Substituting the function g into the matrix M′, a new matrix M′ can be obtained to detect points of interest.

[0090]

[0091] By traversing and comparing, we find and connect the points of interest that meet the requirements. Using the color enhancement detection algorithm to obtain all points of interest, we first find a starting point among all the points of interest. This starting point should be located at the bottom right of all the points of interest. Next, we search for the point of interest with the smallest angle in a certain direction. We connect the starting point to this point of interest. Finally, we repeat the above process.

[0092] Next, the location of the query point of interest determines whether to terminate the search for the point of interest. If the location of the currently found point of interest is the same as the initial point of interest, the calculation terminates and a minimum convex hull is obtained. Otherwise, the search for a qualifying point of interest continues until the initial point of interest is found and connected with a straight line. The above calculations produce a convex hull model, which can be used to determine the detection area of the candidate object.

[0093] Step 3: Perform multi-layer cellular automaton fusion to form the final saliency map and obtain the saliency target.

[0094] Finally, using a multi-layer cellular automaton, each pixel in the saliency map is a cell. In an M-layer automaton, a cell in any saliency map has M-1 neighbors, each located at the same position in other saliency maps. The saliency value of a pixel or cell i represents its probability of being a foreground F: P(i∈F) = S i , then the probability of belonging to background B can be expressed as P(i∈B)=1-S i Represented by. Extract the adaptive threshold of each image, using γ m Represents the binarization threshold of the mth saliency map and performs binarization processing on it. If the saliency value S of cell i i ≥γ m , it is marked as foreground and η i = +1. On the contrary, it means that cell i is background. If cell i is marked as foreground, then the probability that its neighbor j at the same position in other saliency maps is marked as foreground is λ = P(η i = +1|i∈F), where η i is the number of neighbor pixels i marked as foreground, F is the probability that neighbor pixel i is marked as foreground. Similarly, we can use μ=P(η) to represent the probability that neighbor j is background when cell i marks background. i = +1|i∈B), where η i is the number of neighbor pixels i marked as background, and B is the number of neighbor pixels i marked as background. Assuming that λ and μ are a pair of equal constants, the posterior probability P(i∈F|η i =+1) can be expressed by the following formula:

[0095] P(i∈F|η i =+1)∝P(i∈F)P(η i = +1|i∈F) = S i *λ

[0096] The prior probability ratio is defined as Λ(i∈F), and the calculation formula is as follows:

[0097]

[0098] Then the posterior probability ratio Λ(i∈F|η i =+1) can be expressed as follows:

[0099]

[0100] Transform the above formula l=ln(Λ) into the following form:

[0101]

[0102] The prior probability ratio and the posterior probability ratio are expressed as follows:

[0103]

[0104] in Represents the significant value of cell i at time t. Synchronous update mechanism f:S M-1 →S is defined as follows:

[0105]

[0106] in, Represents the significance values of all cells on the mth saliency map at time t. Matrix 1 is a matrix with N elements [1,…,1] T If a cell's neighbor is judged as the foreground, its own significance value will be increased accordingly, that is, Then λ>0.5. According to experience, After N2 updates, the final saliency map can be obtained by the following formula, which shows the salient objects.

[0107]

[0108] Experiments have shown that Figure 2 As shown in the figure, the salient target detection method based on superpixel segmentation can obtain the complete salient target area and further refine the boundary information. The comprehensive results of various evaluation indicators show that it can achieve good performance.

[0109] In summary, the embodiment of the present invention first adopts a superpixel segmentation algorithm to segment the image into superpixels. Secondly, the candidate target area is obtained based on the convex hull model. If the salient target is at the background boundary, the boundary superpixel area will be removed, so that the superpixels around the remaining image are obtained as the background area. This patent proposes an improved weighted multi-layer cellular automaton saliency map fusion method, which performs consistency fusion optimization on the obtained superpixel segmentation saliency map to obtain the final saliency map. Experiments have shown that the salient target detection method based on superpixel segmentation can obtain a complete salient target area and further refine the boundary information. The comprehensive results of various evaluation indicators show that good performance can be achieved.

[0110] If the above functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0111] It will be understood by those skilled in the art that the embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention may be implemented in various computer languages, for example, the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0112] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0113] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0114] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0115] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0116] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A salient object detection method based on superpixel segmentation, characterized in that: The following steps are involved: Acquire the image to be tested; Based on the image to be tested, superpixel segmentation is performed using a superpixel segmentation algorithm to obtain a segmented image; based on the segmented image, a color enhancement detection algorithm is used to process the image to obtain an initial saliency map containing interest points; Based on the initial saliency map containing the interest points, a convex hull model is used to process the initial saliency map to obtain an initial saliency map containing the candidate target detection area; A multi-layer cellular automation is used to discriminate the foreground and background in the initial saliency map containing the candidate target detection area, generate a final saliency map, and obtain a salient target.

2. The method for detecting salient objects based on superpixel segmentation according to claim 1, wherein: The step of obtaining the segmented image comprises: 1) Initialization parameters, including the number of superpixels K, the number of pixels N and the cluster center C k , where k is the number of cluster centers; 2) With the currently set cluster center C k The first neighborhood is set as the center, and the spatial position with the minimum gradient amplitude value is found in the first neighborhood by calculating the image gradient amplitude value as the new cluster center; 3) Based on the new cluster center, a second neighborhood is set, and all pixels in the second neighborhood are classified using a nearest neighbor classification method to obtain a classification result; 4) Repeat steps 2)-3) to iteratively update the cluster centers and classification results until the iteration ends to obtain a final classification result, which is used as the segmented image.

3. The method for detecting salient objects based on superpixel segmentation according to claim 2, wherein: The calculation expression of the image gradient amplitude value is: G(x,y)=||I(x+1,y)-I(X-1,y)|| 2 +||I(x,y+1)-I(x,y-1)|| 2 Where G(x,y) is the image gradient amplitude, and I(x,y) is the color vector of the color feature of the pixel at the (x,y) position.

4. The method for detecting salient objects based on superpixel segmentation according to claim 2, wherein: The steps of classifying by the nearest neighbor classification method include: In the second neighborhood, for each pixel, find the new cluster center with the closest distance, and use the category of the new cluster center as the pixel label to complete the pixel classification process. The distance calculation expression is: Where, d lab and d xy Represents the color feature distance and spatial position feature distance respectively, D s It is d lab and d xy It is obtained by linear superposition, which represents the feature distance. m represents a variable that controls the compactness between superpixels. l, a, and b represent different channels in the color feature space, and (x, y) represent pixel positions.

5. The method for detecting salient objects based on superpixel segmentation according to claim 1, wherein: The step of obtaining an initial saliency map containing interest points comprises: Based on the segmented image, a color space conversion is performed to obtain a converted segmented image; based on the converted segmented image, a color saliency enhancement is performed through a conversion function g to generate a saliency matrix M ′ , where the significance matrix M ′ The expression is: Among them, the conversion function g satisfies: I(v)=-log(p(v)) Where ω(x,y) is the Gaussian weight function, g(I x ) and g(I y ) is the gradient component after color enhancement, P(v) is the probability of event v occurring; Based on the saliency matrix M ′ Calculate the local maximum, filter out all interest points, and remove interest points at the edge to obtain the initial saliency map containing interest points.

6. The method for detecting salient objects based on superpixel segmentation according to claim 1, wherein: The step of obtaining an initial saliency map including the candidate target detection area includes: selecting the starting position of the interest point as the starting point of the convex hull based on the initial saliency map including the interest point, wherein the starting point is located at the lower rightmost position of all the interest points; Taking the starting point of the convex hull as a reference, select the interest points with the smallest angle value according to the preset direction and connect them. Then repeat this step with the selected interest point as the next starting point to form a convex hull area as the candidate target detection area, and obtain an initial saliency map containing the candidate target detection area.

7. The method for detecting salient objects based on superpixel segmentation according to claim 1, wherein: The step of generating the final saliency map comprises: Each pixel in the initial saliency map containing the candidate target detection area is taken as a cell i, and the initial saliency map containing the candidate target detection area is binarized to obtain a foreground feature saliency map and a background feature saliency map; Calculating the prior probability ratio and the posterior probability ratio of each cell i on the foreground feature saliency map and the background feature saliency map; Based on the prior probability ratio and the posterior probability ratio, a synchronous update mechanism is used to synchronously update each cell i to adjust the significance value, and this step is repeated for iterative update to obtain the final updated significance value; Based on the final updated saliency values of the foreground feature saliency map and the background feature saliency map, a calculation is performed to obtain a final saliency map, wherein the final saliency map is expressed as: Where, is the final saliency map, N2 is the number of iterative updates, and m is the mth saliency map.

8. The method for detecting salient objects based on superpixel segmentation according to claim 7, wherein: The step of obtaining a foreground feature saliency map and a background feature saliency map comprises: Calculate the saliency value S of each cell i in the initial saliency map containing the candidate target detection area i ; Set the binarization threshold γ m ; Determine each cell i The significance value S i Is it greater than or equal to γ? m If yes, then cell i is marked as foreground, if no, then cell i is marked as background, and finally a foreground feature saliency map and a background feature saliency map are formed; Among them, if cell i is marked as foreground, then the probability λ of its neighbor j at the same position on the remaining initial saliency map containing the candidate target detection area being marked as foreground satisfies: λ=P(η i =+1|i∈F) Where P is the probability function, η i represents the number of neighbor pixels of i that are marked as foreground, F is the number of neighbor pixels i marked as foreground; If cell i is marked as background, then the probability μ of neighbor j at the same position on the remaining initial saliency map containing the candidate target detection area being marked as background satisfies: μ=P(η i =+1|i∈B); Where η i represents the number of neighbor pixels i marked as background, and B represents the number of neighbor pixels i marked as background.

9. The method for detecting salient objects based on superpixel segmentation according to claim 7, wherein: The expressions of the prior probability ratio and the posterior probability ratio are respectively: Among them, the l=ln(Λ) transformation is performed on equations (1) and (2) to obtain the final expressions of the prior probability ratio and the posterior probability ratio: Where Λ(i∈F) is the prior probability ratio, Λ(i∈F|η i =+1) is the posterior probability ratio, λ is the probability that if cell i is marked as foreground, then its neighbor j at the same position on the remaining initial saliency map containing the candidate target detection area is marked as foreground, and μ is the probability that if cell i is marked as background, then its neighbor j at the same position on the remaining initial saliency map containing the candidate target detection area is marked as background.

10. The method for detecting salient objects based on superpixel segmentation according to claim 7, wherein: The synchronous update mechanism is expressed as: Where, for, Represents the significance values of all cells on the mth saliency map at time t. Matrix 1 is a matrix with N elements [1,…,1] T , is the significance value of all cells on the kth saliency map at time t, γ k is the binarization threshold of the kth image, λ is the probability that cell i is marked as foreground, and its neighbor j at the same position on the remaining initial saliency map containing the candidate target detection area is marked as foreground.

Images