A method for extracting visual saliency of images

Abstract

This invention provides a method for extracting visual saliency of an image, including acquiring an image, performing superpixel segmentation on the acquired image to obtain a superpixel set; representing the constructed hypergraph using an association matrix; extracting features from the image input through a deep learning network; calculating the features of the superpixels; calculating the saliency values ​​of the superpixels; calculating the saliency values ​​of the hyperedges; updating the saliency values ​​of the superpixels associated with the hyperedges; recovering the saliency value of each pixel in the entire image based on the position coordinates of pixels in all superpixels; and mapping the importance of the feature space to the image physical space to obtain the final saliency map of the image. This invention re-clusters the superpixels and uses the hypergraph to represent the superpixels and clustering results. The network structure has two branches after the two fully connected layers: one branch obtains the importance of feature channels and weights the features; the other branch obtains the importance of the feature space.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method for extracting visual saliency of an image. Background Technology

[0002] When faced with a scene, humans automatically process regions of interest while selectively ignoring regions of uninterest. These regions of interest are called saliency regions. The human visual system possesses the ability to quickly search for and locate targets of interest in natural scenes; this visual attention mechanism is crucial for processing visual information in daily life. With the proliferation of massive amounts of data brought about by the internet, how to quickly extract important information from vast amounts of image and video data has become a key issue in the field of computer vision. Visual saliency detection models are the process of predicting which information in an image or video will attract more visual attention through computer vision algorithms. Introducing this visual attention mechanism, namely visual saliency, into computer vision tasks can bring a series of significant benefits and improvements to visual information processing tasks. The advantages of introducing visual saliency are mainly reflected in two aspects: first, it can allocate limited computational resources to more important information in images and videos; second, the results of introducing visual saliency are more in line with human visual cognitive needs. Visual saliency detection has important application value in target recognition, image and video compression, image retrieval, and image retargeting.

[0003] Visual saliency includes two mechanisms: bottom-up and top-down. Bottom-up can be considered data-driven, meaning the image itself attracts the viewer, while top-down relies on conscious attention to the image. Traditional saliency detection models, whether bottom-up or top-down, mostly rely on texture cues within the image, using handcrafted low-level features and heuristics. Traditional saliency detection methods lack the ability to capture semantic information, making them difficult to extend to other complex applications. In recent years, with the development of deep learning technology, more and more researchers have begun using it for saliency detection. Deep learning methods train neural network models to learn more accurate and effective saliency features, thus achieving more accurate and robust saliency detection. Common deep learning methods include: methods based on fully convolutional networks, methods based on recurrent neural networks, and methods based on attention mechanisms. Attention mechanisms simulate the human visual system, focusing attention on important regions in an image and generating a corresponding saliency map by calculating the importance of each pixel. While deep learning-based saliency detection methods have achieved good results, there is still room for improvement.

[0004] In 2017, Shandong University published Hao Jing's master's thesis, "Research on Saliency Detection Method Based on Multi-Scale Hypergraph of Adaptive Image Scaling." The thesis stated that saliency detection is an important research topic in image processing and computer vision. This paper explores a method for saliency detection using multi-scale hypergraphs that adaptively scale images with independent channels. The work done is of great significance to the development of image processing and computer vision. Currently, there are many classic saliency detection methods, but almost none of them consider the impact of differences in the R, G, and B channel values ​​of an image on the saliency detection results. This paper introduces the differences in the R, G, and B channel values ​​of an image into a saliency detection method based on hypergraph modeling, proposing a method for saliency detection using adaptive scaling with independent channels. The specific research content includes: (1) Considering the difference in sensitivity of the human eye to the three primary colors R, G, and B, the pixel values ​​of the image are independently adaptively scaled to obtain the independently adaptively scaled image of the initial image; (2) The method of independently adaptively scaling the pixel values ​​is introduced into the construction of hypervertices and hyperedges, and the adaptive scaling factor is combined with the empirical value of the fixed scale to obtain the adaptive multi-scale hyperedges and hypervertices, so as to design an adaptive multi-scale hypermap based on the independently adaptively scaled image; (3) The designed adaptive multi-scale hypermap based on the independently adaptively scaled image is introduced into the saliency detection method, and a saliency detection method based on the independently adaptively scaled image hypermap is designed, and the effectiveness of the designed saliency detection method is demonstrated by using an image segmentation example. First, the saliency of hypervertices and hyperedges based on independent channel adaptive scaling of the image is calculated using gradient maps to obtain the saliency map of the multi-scale hypermap. Then, the saliency maps of the multi-scale hypermap are fused to obtain the final saliency map of the image hypermap based on independent channel adaptive scaling. Finally, the designed saliency detection method is applied to an image segmentation example to illustrate its effectiveness. (4) The designed saliency detection method based on independent channel adaptive scaling of the image hypermap was tested on the publicly available image sets MSRA-1000, SOD, SED and ImgSal-5, and compared with the previous six classic saliency detection methods. A large number of experiments show that the proposed method improves the saliency detection effect to a certain extent for images with narrow differences in the R, G, and B channel values. The saliency detection method based on independent channel adaptive scaling of the image proposed in this paper can be applied to many image processing and computer vision applications such as image segmentation, target recognition, adaptive image compression, and content-based image retrieval, which helps to improve its performance and efficiency. Although this article uses the hypergraph method to model the image, the saliency detection method still uses the traditional calculation method, without using deep learning methods, and does not realize the mapping between feature saliency and image physical space. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes a method for extracting visual saliency of images. This method re-clusters superpixels and uses a supergraph to represent the association between superpixels and clustering results. In terms of network structure, there are two branches after two fully connected layers. One branch can obtain the importance of feature channels and weight the features; the other branch can obtain the importance of the feature space.

[0006] This invention provides a method for extracting visual saliency from an image, comprising acquiring an image and the following steps:

[0007] Step 1: Perform superpixel segmentation on the image to obtain a superpixel set {a n}, 1≤n≤N, where N is the number of superpixels;

[0008] Step 2: Using the correlation matrix K∈R N×M The constructed hypergraph is represented, where N represents the number of superpixels and M represents the number of hyperedges;

[0009] Step 3: Extract features from the input image using a deep learning network;

[0010] Step 4: Calculate the features of the superpixels;

[0011] Step 5: Calculate the saliency value of the superpixel;

[0012] Step 6: Calculate the significance value of the hyperedge;

[0013] Step 7: Update the saliency values ​​of the superpixels associated with the superedge;

[0014] Step 8: Based on the position coordinates of pixels in all superpixels, recover the saliency value of each pixel in the entire image, and represent the corresponding saliency map as X;

[0015] Step 9: Map the importance of the feature space to the physical space of the image to obtain the final saliency map of the image.

[0016] Preferably, step 1 further includes processing the superpixel set {a} n Clustering is performed, and superpixels of the same class are connected by a superedge to create a supergraph.

[0017] Preferably, in any of the above schemes, the structure of the hypergraph is defined as follows: Among them, v n Indicates superpixel a n The corresponding vertex, v n ∈V, E represents the set of superpixels {a n The set of all hyperedges after clustering, where the m-th element is denoted as e. m∈E, 1≤m≤M, W represents the saliency matrix of the hyperedges in the hypergraph, and the elements are denoted as ω(e i ), which can be initialized as a 1×M zero matrix.

[0018] In any of the above schemes, it is preferred that the element k(v) in the correlation matrix... n ,e m ) represents

[0019]

[0020] In any of the above schemes, it is preferred that the deep learning network has a dual-branch structure, where F represents the original feature map and its shape is (Q, C, H, W), where Q is the batch size, C is the number of channels, and H and W are the height and width of the feature map, respectively.

[0021] In any of the above schemes, it is preferred that the first branch of the deep learning network performs global average pooling on each channel of the feature map to obtain a feature vector with a feature dimension equal to the number of channels, calculated as follows:

[0022]

[0023] Where ⊙ is the dot product symbol, g2 and g1 are the weights of the two fully connected layers, σ is the sigmoid function, G() is the global average pooling function, and F1 is the weighted feature map.

[0024] In any of the above schemes, it is preferred that the second branch of the deep learning network performs global average pooling on all feature channels of the second fully connected layer, and the calculation method is as follows:

[0025]

[0026] F1 and F2 are fused to obtain the final feature map F3;

[0027] F3 = h(F1, F2)

[0028] Where h() is the feature fusion function.

[0029] Preferably, in any of the above solutions, the superpixel a n Features The calculation formula is

[0030]

[0031] Among them, |a n | represents superpixel a n The number of pixels in the superpixel, p(j) represents the j-th element in the superpixel corresponding to the p(j)-th pixel in the image, and F3(p(j)) represents the feature of pixel p(j).

[0032] In any of the above schemes, the preferred option is superpixel a. n The method for calculating the significance value is as follows:

[0033]

[0034] Among them, dist(a n ,a j ) represents superpixel a n and a j distance,

[0035] Obtain the saliency matrix S of all superpixels.

[0036]

[0037] In any of the above schemes, the preferred method for calculating the saliency value of the hyperedge is as follows:

[0038]

[0039] in, It is a matrix multiplication operation. count_ones(K) means to count the number of 1s in each column of matrix K.

[0040] In any of the above schemes, it is preferable that S2 represents the saliency value of all superpixels.

[0041]

[0042] Among them, W T This represents the transpose of W.

[0043] In any of the above schemes, the preferred option is that the formula for the final saliency map is:

[0044]

[0045] Where σ is the sigmoid function, f 1×1 This is a convolution operation with a 1x1 kernel.

[0046] This invention proposes a method for extracting visual saliency from images. Through superpixel segmentation and superpixel re-clustering, it avoids the problem of excessive workload in extracting features from individual image pixels and improves the consistency of object region saliency. In terms of network structure, two branches follow each fully connected layer: one branch obtains the importance of feature channels, and the other branch obtains the importance of the feature space. This achieves a mapping from the importance of the feature space to the physical space, enhancing the image saliency extraction effect. Attached Figure Description

[0047] Figure 1 This is a flowchart of a preferred embodiment of the image visual saliency extraction method according to the present invention.

[0048] Figure 2 This is a schematic diagram of the segmentation result of an embodiment of the superpixel segmentation method of the image visual saliency extraction method according to the present invention.

[0049] Figure 3 An example diagram of a simplified embodiment of the original image for the image visual saliency extraction method according to the present invention.

[0050] Figure 4 This is an example diagram of an embodiment of superpixel segmentation according to the image visual saliency extraction method of the present invention.

[0051] Figure 5 This is a schematic diagram of an embodiment of superpixel-based clustering for the image visual saliency extraction method according to the present invention. Detailed Implementation

[0052] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0053] Example 1

[0054] like Figure 1 As shown, step 100 is performed to obtain the image from which visual saliency needs to be extracted.

[0055] Execute step 110 to perform superpixel segmentation on the image to obtain a superpixel set {a n}, 1≤n≤N, where N is the number of superpixels;

[0056] For the superpixel set {a n Clustering is performed, connecting superpixels of the same class together with a hyperedge to construct a hypergraph. The structure of the hypergraph is defined as follows: Among them, v n v represents the vertex corresponding to the superpixel. n ∈V, E represents the set of superpixels {a n The set of all hyperedges after clustering, where the m-th element is denoted as e. m ∈E, 1≤m≤M, W represents the saliency matrix of the hyperedges in the hypergraph, and the elements are denoted as ω(e i ), which can be initialized as a 1×M zero matrix.

[0057] Perform step 120, with the correlation matrix K∈R N×M The constructed hypergraph is represented, where N represents the number of superpixels and M represents the number of hyperedges. The element k(v) in the correlation matrix... n ,e m ) represents

[0058]

[0059] Step 130 involves extracting features from the input image using a deep learning network. The deep learning network has a dual-branch structure, where F represents the original feature map with a shape of (Q, C, H, W), where Q is the batch size, C is the number of channels, and H and W are the height and width of the feature map, respectively.

[0060] The first branch of the deep learning network performs global average pooling on each channel of the feature map to obtain a feature vector with a feature dimension equal to the number of channels. The calculation formula is as follows:

[0061]

[0062] Where ⊙ is the dot product symbol, g2 and g1 are the weights of the two fully connected layers, σ is the sigmoid function, G() is the global average pooling function, and F1 is the weighted feature map.

[0063] The second branch of the deep learning network performs global average pooling on all feature channels of the second fully connected layer, and the calculation method is as follows:

[0064]

[0065] F1 and F2 are fused to obtain the final feature map F3;

[0066] F3 = h(F1, F2)

[0067] Where h() is the feature fusion function.

[0068] Execute step 140 to calculate the features of the superpixels. The calculation formula is as follows:

[0069]

[0070] Among them, |a n | represents superpixel a n The number of pixels in the superpixel, p(j) represents the j-th element in the superpixel corresponding to the p(j)-th pixel in the image, and F3(p(j)) represents the feature of pixel p(j).

[0071] Perform step 150 to calculate the saliency value of the superpixel. The calculation method is as follows:

[0072]

[0073] Among them, dist(a n ,a j ) represents superpixel a n and a jdistance,

[0074] Obtain the saliency matrix S of all superpixels.

[0075]

[0076] Perform step 160 to calculate the saliency value of the hyperedge. The calculation method is as follows:

[0077]

[0078] in, This is a matrix multiplication operation. `count_ones(K)` represents the number of 1s in each column of matrix K. Step 170 is executed to update the saliency values ​​of the superpixels associated with the superedges. `S2` represents the saliency values ​​of all superpixels.

[0079]

[0080] Among them, W T This represents the transpose of W.

[0081] Execute step 180, recover the saliency value of each pixel in the entire image based on the position coordinates of the pixels in all superpixels, and represent the corresponding saliency map as X.

[0082] Step 190 maps the importance of the feature space to the physical space of the image, yielding the final saliency map of the image, as shown in the formula.

[0083]

[0084] Where σ is the sigmoid function, f 1×1 This is a convolution operation with a 1x1 kernel.

[0085] Example 2

[0086] To reduce the excessive workload of feature extraction from image pixels, a common approach is to perform superpixel segmentation. However, this often results in the same object region being segmented into multiple distinct superpixel regions, leading to uneven saliency in the subsequently generated saliency maps. Therefore, this invention re-clusters the superpixels and uses a supermap to represent the association between the superpixels and the clustering results. Furthermore, an association matrix is ​​used to balance the saliency values ​​of superpixels within the same class, improving the consistency of object region saliency. In terms of network structure, two branches follow the two fully connected layers. One branch obtains the importance of feature channels and weights the features; the other branch obtains the importance of the feature space, allowing for both feature space-based weighting and mapping of feature space importance to the importance in the image's physical space. Through these key processing steps, this invention achieves better image saliency extraction results.

[0087] A method for extracting visual saliency of an image includes the following steps:

[0088] (1) Perform superpixel segmentation on the input image / to obtain the superpixel set {a n}, 1≤n≤N, where N is the number of superpixels. For the superpixel set {a n Clustering is performed, connecting superpixels of the same class together with a hyperedge to construct a hypergraph. The hypergraph structure is defined as follows: Among them, v n Indicates superpixel a n The corresponding vertex, v n Let E ∈ V, and let E be the set of all hyperedges, with the m-th element in the set denoted as e. m ∈E, 1≤m≤M, W represents the saliency matrix of the hyperedges in the hypergraph, and the elements are denoted as ω(e i ), which can be initialized as a 1×M zero matrix.

[0089] (2) Using the correlation matrix K∈R N×M The constructed hypergraph is represented as follows: N represents the number of superpixels, M represents the number of hyperedges, and k(v) is an element in the incidence matrix. n ,e m ) represents

[0090]

[0091] (3) The image input is processed by a deep learning network for feature extraction.

[0092] Deep learning networks have a two-branch structure. F represents the original feature map, with a shape of (Q, C, H, W), where Q is the batch size, C is the number of channels, and H and W are the height and width of the feature map, respectively.

[0093] One branch performs global average pooling on each channel of the feature map, resulting in a feature vector with a feature dimension equal to the number of channels. The calculation formula is as follows:

[0094]

[0095] Where ⊙ is the dot product symbol, g2 and g1 are the weights of the two fully connected layers, σ is the sigmoid function, G() is the global average pooling function, and F1 is the weighted feature map.

[0096] Another branch performs global average pooling on all feature channels of the second fully connected layer, calculated as follows:

[0097]

[0098] F1 and F2 are fused to obtain the final feature map F3.

[0099] F3 = h(F1, F2)

[0100] h() is the feature fusion function.

[0101] (4) Calculate the features of the superpixels using the following formula:

[0102]

[0103] |a n | represents superpixel a n The number of pixels in the superpixel, p(j) represents the j-th element in the superpixel corresponding to the p(j)-th pixel in the image, and F3(p(j)) represents the feature of pixel p(j).

[0104] (5) Calculate the saliency value of the superpixel. Superpixel a n The method for calculating the significance value is as follows:

[0105]

[0106] Among them, dist(a n ,a j ) represents superpixel a n and a j The distance. Obtain the saliency matrix S of all superpixels.

[0107]

[0108] (6) Calculate the significance value of the hyperedge. The method for calculating the significance value of the hyperedge is as follows:

[0109]

[0110] It is a matrix multiplication operation. count_ones(K) means to count the number of 1s in each column of matrix K.

[0111] (7) Update the saliency values ​​of the superpixels associated with the superedges. S2 represents the saliency values ​​of all superpixels.

[0112]

[0113] W T This represents the transpose of W.

[0114] (8) Based on the position coordinates of the pixels in all superpixels, recover the saliency value of each pixel in the whole image and represent the corresponding saliency map as X.

[0115] (9) Map the importance of the feature space to the physical space of the image to obtain the final saliency map of the image.

[0116]

[0117] σ is the sigmoid function, f 1×1 (X) represents the convolution operation performed on the features with a kernel size of 1x1.

[0118] This invention achieves:

[0119] (1) By superpixel segmentation and superpixel re-clustering, the problem of too much workload in extracting features from image pixels is avoided on the one hand, and the consistency of object region saliency is improved on the other hand.

[0120] (2) In terms of network structure, there are two branches after the two fully connected layers. One branch can obtain the importance of the feature channel, and the other branch can obtain the importance of the feature space.

[0121] (3) Realize the mapping of importance in feature space to physical space to improve the extraction effect of image saliency.

[0122] Example 3

[0123] This example demonstrates the generation of superpixels.

[0124] like Figure 2 As shown, the SLIC superpixel segmentation method is used, and different numbers of blocks are used in the segmentation results. Figure 2 The top-left image has 20 segments. Figure 2 The image in the upper right corner is divided into 50 blocks. Figure 2 The bottom left image has 100 segments. Figure 2 The bottom right image has 300 segments.

[0125] Example 4

[0126] This example demonstrates the clustering of superpixels.

[0127] To illustrate the process, we simplify the original image to... Figure 3 The diagram shown illustrates how superpixel segmentation is performed on the simplified original image to obtain 8 superpixels, as shown below. Figure 4 As shown. Clustering the superpixels yields four classes, as shown in the figure. Figure 5 As shown.

[0128] Example 5

[0129] This embodiment demonstrates a method for constructing a hypergraph. In the constructed hypergraph, v represents a vertex (a vertex is a superpixel), and e represents a hyperedge (a hyperedge is a cluster obtained from superpixel clustering). Hyperedge e1 includes two vertices v1, v3, v5, and v7; hyperedge e2 includes two vertices v2, v4, v6, and v8; and hyperedge e3 includes two vertices v9, v1, v2, v3, v4, v5, and v7. 11 v 13 v 15 Four vertices, hyperedge e4 includes v 10 v 12 v 14 v 16 Two vertices. Figure 4 The hypergraph is represented by an incidence matrix. If vertex v is in hyperedge e, the element in the matrix is ​​1; otherwise, it is 0.

[0130] Example 6

[0131] This embodiment demonstrates a deep learning network structure.

[0132] Input layer: The network receives the input image.

[0133] Convolutional layer 1: The first convolutional layer, followed by an activation function, which can be ReLU.

[0134] Convolutional layer 2: The second convolutional layer, followed by an activation function.

[0135] Convolutional layer 3: The third convolutional layer, which continues to be connected to the activation function.

[0136] Pooling layer (optional): After the convolutional layer, a pooling layer can be used to reduce the spatial dimension of the feature map.

[0137] Fully connected layer 1: The feature map output by the convolutional layer will be flattened and passed through the first fully connected layer.

[0138] Fully connected layer 2: Then through a second fully connected layer.

[0139] Branch 1: Uses global average pooling to aggregate spatial information and learns channel importance weights through one or more fully connected layers. The channel importance weights are normalized using the sigmoid function, and the channel importance weights are multiplied by the corresponding channel features to obtain weighted channel features.

[0140] Branch 2: Starting from the output of fully connected layer 2, feature space importance may be learned through additional layers, for example, by using the SE module in SENet.

[0141] Feature fusion: The output features of branch 1 and branch 2 are fused, possibly through stacking, weighted averaging or more complex fusion strategies.

[0142] Output layer: The fused features are passed through the final output layer for classification or prediction of other tasks.

[0143] Example 7

[0144] This example demonstrates the calculation results of the significance value.

[0145] The saliency value of each superpixel was calculated, and the results are shown in Table 1.

[0146] Table 1 shows the calculation results of the saliency values ​​for each superpixel.

[0147]

[0148]

[0149] The significance value of each hyperedge was calculated, and the results are shown in Table 2.

[0150] Table 2 shows the calculation results of the significance values ​​for each hyperedge.

[0151] Super Edge e1 e2 e3 e4 Significance 0.28 0.5 0.55 0.18

[0152] Based on the saliency value of the superedge, update the saliency value of the superpixel, as shown in Table 3.

[0153] Table 3 shows the calculation results of the significance values ​​of the updated superpixels.

[0154] Superpixel v1 v2 v3 v4 v5 v6 v7 v8 v9 v10 v11 v12 v13 v14 v15 v16 Significance 0.28 0.5 028 0.5 0.28 0.5 0.28 0.5 0.55 0.18 0.55 0.18 0.55 0.18 0.55 0.18

[0155] To better understand this invention, specific embodiments have been described in detail above, but these are not intended to limit the invention. Any simple modifications made to the above embodiments based on the technical essence of this invention still fall within the scope of this invention. Each embodiment in this specification focuses on its differences from other embodiments; similar or identical parts between embodiments can be referred to mutually. For system embodiments, since they basically correspond to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

Claims

1. A method for extracting visual saliency from an image, comprising acquiring an image, characterized in that, It also includes the following steps: Step 1: Perform superpixel segmentation on the image to obtain a superpixel set {a n }, N, where N is the number of superpixels, for the superpixel set {a n Clustering is performed, and superpixels of the same class are connected by a superedge to create a supergraph; Step 2: Using the correlation matrix The constructed hypergraph is represented, where N represents the number of superpixels and M represents the number of hyperedges; Step 3: The input image is processed through a deep learning network for feature extraction. This deep learning network has a dual-branch structure, where 𝐹 represents the original feature map, and its shape is... Where Q is the batch size, It is the number of channels. and These are the height and width of the feature map, respectively. The first branch of the deep learning network performs global average pooling on each channel of the feature map to obtain a feature vector with a feature dimension equal to the number of channels. The calculation formula is as follows: ， Here, ⊙ represents the dot product symbol, and g2 and g1 are the weights of the two fully connected layers, respectively. It is the sigmoid function, and G() is the global average pooling function. It is the weighted feature map; The second branch of the deep learning network performs global average pooling on all feature channels of the second fully connected layer, and the calculation method is as follows: ， Will and The feature maps are then fused to obtain the final feature map. ; ， Where h() is the feature fusion function; Step 4: Calculate the features of the superpixels; Step 5: Calculate the saliency value of the superpixel; Step 6: Calculate the significance value of the hyperedge; Step 7: Update the saliency values ​​of the superpixels associated with the superedge; Step 8: Based on the position coordinates of pixels in all superpixels, recover the saliency value of each pixel in the entire image, and represent the corresponding saliency map as X; Step 9: Map the importance of the feature space to the physical space of the image to obtain the final saliency map of the image. The formula for the final saliency map is as follows: ， Where, f 1×1 This is a convolution operation with a 1x1 kernel.

2. The method for extracting visual saliency of an image as described in claim 1, characterized in that, The structure of the hypergraph is defined as follows: ,in, Superpixels The corresponding vertex, E represents the set of superpixels {a n The set of all hyperedges after clustering, where the m-th element is denoted as . , W represents the saliency matrix of the hyperedges in the hypergraph, and its elements are denoted as . It can be initialized as a 1×M zero matrix.

3. The method for extracting visual saliency of an image as described in claim 2, characterized in that, Elements in the correlation matrix Represented as 4. The method for extracting visual saliency of an image as described in claim 3, characterized in that, The superpixel Features The calculation formula is ， in, Superpixels The number of pixels in the middle, This indicates that the j-th element in the superpixel corresponds to the j-th element in the image. 1 pixel, Represents pixels Its characteristics.

5. The method for extracting visual saliency of an image as described in claim 4, characterized in that, Superpixel The method for calculating the significance value is as follows: ， in, Superpixels and distance, Obtain the saliency matrix S of all superpixels.

6. The method for extracting visual saliency of an image as described in claim 5, characterized in that, The significance value of the hyperedge is calculated as follows: ， in, It is a matrix multiplication operation. count_ones(K) means calculating the number of 1s in each column of matrix K, that is, how many superpixels are associated with each superedge.

Images