A method for salient object detection in remote sensing images

Abstract

The present invention discloses a remote sensing image salient target detection method, the method comprising the following steps: S1, obtaining remote sensing image data including a training set and a test set, and constructing a remote sensing image salient target detection model including a detection feature encoder and a cascade feature decoder; S2, introducing an attention mechanism, a feature flow mechanism and a cascade decoding mechanism, and training the remote sensing image salient target detection model based on the remote sensing image data of the training set, until the training is stopped after the preset loss function converges, and the trained remote sensing image salient target detection model is obtained; S3, using the trained remote sensing image salient target detection model to predict salient targets on the remote sensing image data of the test set, and then outputting the corresponding salient map. The present invention performs feature decoding based on a cascade structure, improves the problem of missed detection and misdetection of small targets in remote sensing images, and improves the prediction confidence of salient areas, and can predict more accurate salient target boundaries.

Description

Technical Field

[0001] The present invention mainly relates to the technical field of remote sensing image application, and in particular to a remote sensing image salient target detection method. Background Art

[0002] With the explosive growth of remote sensing image data, the traditional remote sensing image utilization method of manual visual interpretation can no longer meet the actual needs. Therefore, it is urgent to develop intelligent interpretation methods for remote sensing images. As an important pre-processing step of computer vision, salient object detection has achieved good results in natural scenes. However, due to the many characteristics of remote sensing scenes such as different shooting angles, diverse types of objects, and complex backgrounds, there are still few salient object detection methods for remote sensing images. At the same time, in the process of detecting salient objects in remote sensing images, the existing methods have poor detection effects on the edge areas of salient objects, and are prone to misdetection and missed detection of small objects, which is still a long way from being realized. Summary of the invention

[0003] In view of this, the purpose of the present invention is to provide a method for detecting salient targets in remote sensing images. The method adopts an encoder-decoder structure, introduces an attention mechanism, a feature flow mechanism, and a cascade decoding mechanism, and designs a new loss function to train a target detection model. Then, the salient targets in the remote sensing images are detected through the trained target detection model, which effectively improves the detection effect of the edges of salient targets in remote sensing images, and at the same time improves the situations of missed detection and false detection of small targets.

[0004] The remote sensing image salient target detection method of the present invention comprises the following steps:

[0005] S1. Obtain remote sensing image data including a training set and a test set, and construct a remote sensing image salient object detection model including a detection feature encoder and a cascade feature decoder;

[0006] S2, introducing an attention mechanism, a feature flow mechanism and a cascade decoding mechanism, and training the remote sensing image salient target detection model based on the remote sensing image data of the training set, until the training is stopped after the preset loss function converges, and a trained remote sensing image salient target detection model is obtained;

[0007] S3. Use the trained remote sensing image salient target detection model to predict salient targets on the remote sensing image data of the test set, and then output the corresponding salient map.

[0008] Furthermore, the detection feature encoder is a dense attention flow encoder, which is improved based on the VGG16 network as the backbone network. The improvement process is: remove the last three fully connected layers of the VGG16 network and truncate before the last level of pooling layer of the VGG16 network to obtain the dense attention flow encoder.

[0009] Furthermore, the specific implementation of step S2 includes:

[0010] S21. Introduce the attention mechanism to extract the output features of the last layer of each part from the improved VGG16 network, and merge the output feature dimensions based on the preset spatial pixel relationship matrix to construct an operation matrix between pixels, thereby representing the relationship between each pixel;

[0011] S22, performing normalization processing based on the operation matrix between pixels to obtain attention weights, and multiplying the output features after merging the dimensions by the attention weights to obtain features weighted by spatial self-attention;

[0012] S23, using the residual connection method to add the output features and the features weighted by the spatial attention, and through the connection channel attention mechanism, the deep features of the output are obtained, and the process is expressed by the formula:

[0013] F=CA(f+δ·(f*(Re -1 (Re(f)⊙R))))

[0014] In the formula, Re -1 represents the inverse operation of the output feature merging dimension, R represents the pixel relationship matrix, * represents element-by-element multiplication, δ represents a learnable coefficient, CA(·) represents the channel attention mechanism, and f represents the initial feature output by the backbone network;

[0015] S24, upsampling and 1×1 convolution of the deep feature to adjust the size and channel of the deep feature to be consistent with the current feature;

[0016] S25, based on a preset step-by-step concatenation module, concatenate the deep features after upsampling and 1×1 convolution with the current features in the order from shallow to deep layers from the next layer of the current features;

[0017] S26, adjusting the number of channels of the concatenated features to the number of channels of the deep features output by the detection feature encoder, and inputting the concatenated features into the cascade feature decoder for decoding;

[0018] S27. Use the Sigmoid function to activate the final output of the cascade feature decoder, thereby completing the training of the remote sensing image salient target detection model.

[0019] Furthermore, the preset spatial pixel relationship matrix in step S21 is expressed by the formula:

[0020] M={(Re(f)) T ⊙Re(f)} T

[0021] Where Re() represents the operation of merging the last two dimensions of the output feature into one dimension, ⊙ represents the matrix multiplication operation, and T represents the transpose.

[0022] Furthermore, the normalization process based on the operation matrix between pixels in step S22 is expressed by the formula:

[0023]

[0024] Where r(x, y) represents the importance of the influence of pixel x on pixel y, m(x, y) represents the element in the pixel relationship matrix, and e represents a natural constant.

[0025] Furthermore, it also includes step S23', using a multi-level pyramid to fuse multi-scale spatial attention to extract information from the output features, specifically: updating the output features to three channels with different resolutions through 2x and 4x downsampling, using a multi-level pyramid to fuse multi-scale spatial attention to refine features of different scales, and fusing the refined features with the output features based on the residual structure, and then fusing the three-level features in order from low to high resolution, thereby obtaining the deep features of the multi-level pyramid fused multi-scale spatial attention weights, and finally merging the deep features output by step S23 with the deep features of the multi-level pyramid fused multi-scale spatial attention weights.

[0026] Furthermore, in step S25, the deep features after upsampling and 1×1 convolution are concatenated with the current features, which is expressed as:

[0027] F k =Conc(Conv(Up(f5)),...Conv(Up(F k-1 )), F k )

[0028] In the formula, Up(·) means upsampling to align the deep features with the current features, and F k It represents the k-th level feature fed into the cascade decoder, F5 represents the 5th level feature fed into the cascade decoder, and Conv represents the convolutional layer.

[0029] Furthermore, the preset loss function is a combined loss function with different weight coefficients, which is expressed by the formula:

[0030] L=ω1L P+ω2L R +ω3L MAE +ω4L S

[0031] Where, L P , L R , L MAE and L s They represent the precision loss term, the recall loss term, the mean absolute error loss term, and the structural similarity loss term, respectively. ω1, ω2, ω3, and ω4 represent L P , L R , L MAE and L s The weight coefficient of , where:

[0032]

[0033]

[0034]

[0035] L S =1-S measure

[0036] S measure =α×S o +(1-α)×S r

[0037] Where N is the total number of samples, n is the sample number, j is the pixel number in the height direction of the remote sensing image, i is the pixel number in the width direction of the remote sensing image, ε is the preset constant, W and H are the width and height of the remote sensing image, s(i, J)∈S is the predicted value of each pixel, g(i, j)∈G is the true value of each pixel, S is the significant prediction result, G is the true label, S is the predicted value of each pixel, and G is the true label. r is the region-oriented similarity measure, S o is the similarity measure for object structure, and α represents a hyperparameter, which is used to weigh the similarity measure for region and the similarity measure for object structure.

[0038] Furthermore, the method further includes step S4, comparing the output corresponding saliency map with the true value map, thereby measuring the level of the saliency map generated by the remote sensing image salient target model.

[0039] Furthermore, the specific implementation method of step S4 is: based on the preset indicators PR curve, F value, mean absolute loss and S value, the salient map generated by the remote sensing image salient target model is measured.

[0040] Compared with the prior art, the remote sensing image salient target detection method of the present invention has the following advantages:

[0041] (1) The present invention uses a cascade structure for feature decoding, so that more high-level semantic features can guide the feature decoding process, effectively improving the problem of missed detection and false detection of small targets in remote sensing images.

[0042] (2) The present invention designs a new loss function to train the remote sensing image salient target detection model, which improves the prediction confidence of the salient area and enables the model to predict the salient target boundary more accurately. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The accompanying drawings constituting a part of the present invention are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the accompanying drawings:

[0044] Figure 1 It is a flow chart of a remote sensing image salient target detection method in the present invention;

[0045] Figure 2 It is a schematic diagram of the structure of the remote sensing image salient target detection method in the present invention;

[0046] Figure 3 Schematic diagram of the self-attention mechanism in the present invention;

[0047] Figure 4 is a PR curve result diagram of an embodiment of the present invention;

[0048] Figure 5 The results of improving the false detection and missed detection of small targets are shown in Figure 1, where (a) is the remote sensing image, (b) is the true value map of the salient target, (c) is the original generated salient map, and (d) is the salient map generated by this method.

[0049] Figure 6 The following is the result of improving the prediction effect of the boundary area of ​​salient targets, where (a) is the remote sensing image, (b) is the true value map of the salient target, (c) is the original generated salient map, and (d) is the salient map generated by this method. DETAILED DESCRIPTION

[0050] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0051] See also Figure 1-Figure 6 The remote sensing image salient target detection method of the present invention comprises the following steps:

[0052] S1. Obtain remote sensing image data including a training set and a test set, and construct a remote sensing image salient object detection model including a detection feature encoder and a cascade feature decoder;

[0053] In this step, the detection feature encoder is a dense attention flow encoder, which is improved based on the VGG16 network as the backbone network. The improvement process is: remove the last three fully connected layers of the VGG16 network and truncate before the last pooling layer of the VGG16 network to obtain the dense attention flow encoder. Therefore, the feature dimensions of the first four layers in the improved VGG16 network are W and H are the width and height of the remote sensing image respectively, k is the backbone network layer, and the feature dimension of the last layer is consistent with that of the fourth layer because the pooling layer is removed. At the same time, after each feature extraction layer is completed, the features of the current level are selected and refined and then sent to the next level for extraction.

[0054] In this embodiment, EORSSD is used as the remote sensing image dataset, and 1400 remote sensing images are randomly selected from the remote sensing image dataset as a training set, and 600 remote sensing images are randomly selected as a test set.

[0055] S2, introducing an attention mechanism, a feature flow mechanism and a cascade decoding mechanism, and training the remote sensing image salient target detection model based on the remote sensing image data of the training set, until the training is stopped after the preset loss function converges, and a trained remote sensing image salient target detection model is obtained;

[0056] In this step, it is specifically as follows:

[0057] S21. Introduce the attention mechanism to extract the output features of the last layer of each part from the improved VGG16 network, and merge the output feature dimensions based on the preset spatial pixel relationship matrix to construct an operation matrix between pixels, thereby representing the relationship between each pixel;

[0058] The preset spatial pixel relationship matrix is ​​expressed as follows:

[0059] M={(Re(f)) T ⊙Re(f)} T

[0060] In the formula, Re() represents the operation of merging the last two dimensions of the output feature into one dimension, ⊙ represents the matrix multiplication operation, and T represents the transpose;

[0061] S22, performing normalization processing based on the operation matrix between pixels to obtain the attention weight, and multiplying the output feature after the merged dimension by the attention weight, and then restoring the temperature of the output feature to obtain the feature weighted by the spatial self-attention, and the obtained feature weighted by the spatial self-attention has global information;

[0062] Among them, the normalization processing based on the operation matrix between pixels is expressed by the formula:

[0063]

[0064] Where r(x, y) represents the importance of the influence of pixel x on pixel y, m(x, y) represents the element in the pixel relationship matrix, and e represents a natural constant;

[0065] S23, using the residual connection method to add the output features and the features weighted by the spatial attention, and through the connection channel attention mechanism, the deep features of the output are obtained, and the process is expressed by the formula:

[0066] F=CA(f+δ·(f*(Re -1 (Re(f)⊙R))))

[0067] In the formula, Re -1 represents the inverse operation of the output feature merging dimension, R represents the pixel relationship matrix, * represents element-by-element multiplication, δ represents a learnable coefficient, CA(·) represents the channel attention mechanism, and f represents the initial feature output by the backbone network;

[0068] In this embodiment, after the remote sensing image is subjected to improved VGG16 network feature extraction and feature refinement, five features of different scales are finally formed, and the deeper layers of the five features of different scales contain more semantic features, while the shallower layers retain more detail features.

[0069] S24, upsampling and 1×1 convolution of the deep feature to adjust the size and channel of the deep feature to be consistent with the current feature;

[0070] S25, based on the preset step-by-step splicing module, the deep features after upsampling and 1×1 convolution are spliced ​​with the current features in the order from shallow to deep from the next layer of the current feature. The splicing process is expressed by the formula:

[0071] F k =Conc(Conv(Up(F5)),...Conv(Up(F k-1 )), F k )

[0072] In the formula, Up(·) means upsampling to align the deep features with the current features, and F k represents the k-th level feature fed into the cascade decoder, F5 represents the 5th level feature fed into the cascade decoder, and Conv represents the convolutional layer;

[0073] In this embodiment, in order to achieve complete extraction of image features, the multi-level features are fused by cascading attention from shallow to deep layers. Taking the GCA4 module as an example, the GCA4 module receives the output features of the GCA1 module, the GCA2 module, and the GCA3 module, respectively, and re-adjusts the number of channels to 1, thereby forming the final attention map, which is expressed as follows:

[0074] A4=Conv(Conc(A1,A2,A3,A4))

[0075] The attention map is then multiplied by the refined features, and a residual connection is used to generate deep features that are fed into the cascaded feature decoder.

[0076] S26, adjusting the number of channels of the concatenated features to the number of channels of the deep features output by the detection feature encoder, and inputting the concatenated features into the cascade feature decoder for decoding;

[0077] S27. Use the Sigmoid function to activate the final output of the cascade feature decoder, thereby completing the training of the remote sensing image salient target detection model.

[0078] In this embodiment, since the deep features have the richest semantic features, they can guide each level of decoder to decode; however, all features from deeper layers have more semantic information than features from shallow layers. By using the cascade feature decoder, in addition to the deepest global features, the deep features obtained by each level of encoder can guide the shallow decoder, thereby contributing to the generation of the final saliency map. Each decoder unit receives the output of the previous decoder and the concatenated features of the deep layer, and uses the Sigmoid function to activate the output of the last decoder to obtain the final predicted saliency map.

[0079] S3. Use the trained remote sensing image salient target detection model to predict salient targets on the remote sensing image data of the test set, and then output the corresponding salient map.

[0080] In this embodiment, by performing salient target prediction on the remote sensing image data of the test set based on the trained remote sensing image salient target detection model, a more accurate salient map of the salient target boundary can be obtained.

[0081] The preset loss function is a combined loss function with different weight coefficients, which is expressed as follows:

[0082] L=ω1L P +ω2L R +ω3L MAE +ω4L S

[0083] Where, L P , L R , L MAE and L s They represent the precision loss term, the recall loss term, the mean absolute error loss term, and the structural similarity loss term, respectively. ω1, ω2, ω3, and ω4 represent L P , L R , L MAE and L s The weight coefficient of , where:

[0084]

[0085]

[0086]

[0087] L S =1-S measure

[0088] S measure =α×S o +(1-α)×S r

[0089] Where N is the total number of samples, n is the sample number, j is the pixel number in the height direction of the remote sensing image, i is the pixel number in the width direction of the remote sensing image, ε is the preset constant, W and H are the width and height of the remote sensing image, s(i, j)∈S is the predicted value of each pixel, g(i, j)∈G is the true value of each pixel, S is the significant prediction result, G is the true label, S is the predicted value of each pixel, and G is the true label. r is the region-oriented similarity measure, S o is the similarity measure for object structure, and α represents a hyperparameter, which is used to weigh the similarity measure for region and the similarity measure for object structure.

[0090] In this embodiment, since the difference in images is obtained by comparing the structural information between images using structural similarity, which is more in line with the perception of the human eye, a combined loss function with different weight coefficients is used as the preset loss function, which can solve the problem that the cross entropy loss function has poor detection ability for edge parts during the detection of significant targets.

[0091] In another embodiment, step S4 is also included, comparing the output corresponding saliency map with the true value map, thereby measuring the level of the saliency map generated by the remote sensing image salient target model, specifically: measuring the saliency map generated by the remote sensing image salient target model based on preset indicators PR curve, F value, mean absolute loss and S value.

[0092] In this embodiment, the saliency map generated by the model is compared with the true value map to quantitatively measure the level of saliency map generation. Specifically, four indicators, PR curve, F value, mean absolute loss (MAE) and S value, are used for evaluation.

[0093] Among them, Precision refers to the proportion of correctly predicted positive samples to all predicted positive samples, that is, the precision rate; Recall refers to the proportion of correctly predicted positive samples to positive samples in the label, that is, the recall rate. By adjusting the threshold between (0,1), all (Precision, Recall) values ​​can be obtained, and the Precision-Recall (PR) curve can be obtained by connecting them in sequence. Therefore, the closer the PR curve is to the (1,1) point on the coordinate axis, the better the performance of the model; Figure 4 As shown, Figure 4 The PR curve of the remote sensing image salient target detection method in this embodiment is shown.

[0094] The definition of F value is

[0095]

[0096] In the formula, β 2 Set to 0.3 to emphasize the importance of Precision;

[0097] The mean absolute loss (MAE) is an indicator to measure the absolute error between the significant prediction map and the true value map, which is expressed as follows:

[0098]

[0099] In the formula, S represents the significant prediction result and G represents the true label.

[0100] S measure The value is an indicator that measures the generated saliency map based on the degree of structural similarity, which is expressed as:

[0101] S measure =α×S o +(1-α)×S r

[0102] In the formula, S r is the region-oriented similarity measure, S o is the similarity measure for object structures, and α represents a hyperparameter.

[0103] In this embodiment, the test set is subjected to salient target detection by remote sensing image salient target detection method, and the detection results are shown in Table 1 and Figure 5 , Figure 6 shown.

[0104] Table 1 shows the results of salient target detection in remote sensing images.

[0105] Evaluation indicators F MAE S value 0.9031 0.0048 0.9189

[0106] from Figure 5 and Figure 6 It can be seen that the remote sensing image salient target detection method can not only accurately predict the salient targets in the remote sensing images, but also accurately predict the boundary areas of the salient targets. At the same time, the prediction in the small target scene is also relatively accurate, reducing the cases of missed detection and false detection.

[0107] In another embodiment, step S23' is also included, and information is extracted from the output features by using a multi-level pyramid fusion of multi-scale spatial attention. Specifically, the output features are updated to three channels with different resolutions by downsampling by 2 times and 4 times, and features of different scales are refined by using a multi-level pyramid fusion of multi-scale spatial attention. The refined features are fused with the output features based on the residual structure, and the three-level features are fused in order from low to high resolution to obtain the deep features of the multi-level pyramid fusion of multi-scale spatial attention weights. Finally, the deep features output by step S23 are merged with the deep features of the multi-level pyramid fusion of multi-scale spatial attention weights.

[0108] In this embodiment, in addition to the attention between each pixel, the multi-scale attention of the entire image space can also extract useful information. Specifically, after obtaining the feature output with self-attention, the GCA module will also use a multi-level pyramid to fuse multi-scale spatial attention.

[0109] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for detecting salient targets in remote sensing images, characterized in that: The method comprises the following steps: S1. Obtain remote sensing image data including a training set and a test set, and construct a remote sensing image salient object detection model including a detection feature encoder and a cascade feature decoder; S2, introducing an attention mechanism, a feature flow mechanism and a cascade decoding mechanism, and training the remote sensing image salient target detection model based on the remote sensing image data of the training set, until the training is stopped after the preset loss function converges, and a trained remote sensing image salient target detection model is obtained; S3, using the trained remote sensing image salient target detection model to predict salient targets on the remote sensing image data of the test set, and then outputting a corresponding salient map; The detection feature encoder is a dense attention flow encoder, which is obtained by improving the VGG16 network as the main network. The improvement process is: removing the last three fully connected layers of the VGG16 network and truncating before the last pooling layer of the VGG16 network, thereby obtaining the dense attention flow encoder; The specific implementation of step S2 includes: S21. Introduce the attention mechanism to extract the output features of the last layer of each part from the improved VGG16 network, and merge the output feature dimensions based on the preset spatial pixel relationship matrix to construct an operation matrix between pixels, thereby representing the relationship between each pixel; S22, performing normalization processing based on the operation matrix between pixels to obtain attention weights, and multiplying the output features after merging the dimensions by the attention weights to obtain features weighted by spatial self-attention; S23, using the residual connection method to add the output features and the features weighted by the spatial attention, and through the connection channel attention mechanism, the deep features of the output are obtained, and the process is expressed by the formula: F=CA(f+δ·(f*(Re -1 (Re(f)⊙R)))) In the formula, Re -1 represents the inverse operation of the output feature merging dimension, R represents the pixel relationship matrix, * represents element-by-element multiplication, δ represents a learnable coefficient, CA(·) represents the channel attention mechanism, and f represents the initial feature output by the backbone network; S24, upsampling and 1×1 convolution of the deep feature to adjust the size and channel of the deep feature to be consistent with the current feature; S25, based on a preset step-by-step concatenation module, concatenate the deep features after upsampling and 1×1 convolution with the current features in the order from shallow to deep layers from the next layer of the current features; S26, adjusting the number of channels of the concatenated features to the number of channels of the deep features output by the detection feature encoder, and inputting the concatenated features into the cascade feature decoder for decoding; S27. Use the Sigmoid function to activate the final output of the cascade feature decoder, thereby completing the training of the remote sensing image salient target detection model.

2. The remote sensing image salient target detection method according to claim 1, characterized in that: The preset spatial pixel relationship matrix in step S21 is expressed by the formula: M={(Re(f)) T ⊙Re(f)} T Where Re() represents the operation of merging the last two dimensions of the output feature into one dimension, ⊙ represents the matrix multiplication operation, and T represents the transpose.

3. The remote sensing image salient target detection method according to claim 2, characterized in that: The normalization process based on the operation matrix between pixels in step S22 is expressed by the formula: Where r(x,y) represents the importance of the influence of pixel x on pixel y, m(x,y) represents the element in the pixel relationship matrix, and e represents a natural constant.

4. The remote sensing image salient target detection method according to claim 3, characterized in that: The method also includes step S23', using a multi-level pyramid to fuse multi-scale spatial attention to extract information from the output features, specifically: updating the output features to three channels with different resolutions through 2x and 4x downsampling, using a multi-level pyramid to fuse multi-scale spatial attention to refine features of different scales, and fusing the refined features with the output features based on the residual structure, and then fusing the three-level features in order from low to high resolution to obtain the deep features of the multi-level pyramid fused multi-scale spatial attention weights, and finally merging the deep features output by step S23 with the deep features of the multi-level pyramid fused multi-scale spatial attention weights.

5. The remote sensing image salient target detection method according to claim 4, characterized in that: In step S25, the deep features after upsampling and 1×1 convolution are concatenated with the current features, which can be expressed as: F k =Conc(Conv(Up(F5)),...Conv(Up(F k-1 )),F k ) In the formula, Up(·) means upsampling to align the deep features with the current features, and F k It represents the k-th level feature fed into the cascade decoder, F5 represents the 5th level feature fed into the cascade decoder, and Conv represents the convolutional layer.

6. The remote sensing image salient target detection method according to claim 5, characterized in that: The preset loss function is a combination loss function with different weight coefficients, which is expressed as follows: L=ω1L P +ω2L R +ω3L MAE +ω4L S Where, L P , L R , L MAE and L s They represent the precision loss term, the recall loss term, the mean absolute error loss term, and the structural similarity loss term, respectively. ω1, ω2, ω3, and ω4 represent L P , L R , L MAE and L s The weight coefficient of , where: L S =1-S measure S measure =α×S o +(1-a)×S r Where N is the total number of samples, n is the sample number, j is the pixel number in the height direction of the remote sensing image, i is the pixel number in the width direction of the remote sensing image, ε is the preset constant, W and H are the width and height of the remote sensing image, s(i,j)∈S is the predicted value of each pixel, g(i,j)∈G is the true value of each pixel, S is the significant prediction result, G is the true label, S is the predicted value of each pixel, and G is the true label. r is a region-oriented similarity measure, S o is the similarity measure for object structure, and α represents a hyperparameter, which is used to weigh the similarity measure for region and the similarity measure for object structure.

7. The remote sensing image salient target detection method according to claim 6, characterized in that: The method further includes step S4, comparing the output corresponding saliency map with the true value map, thereby measuring the level of the saliency map generated by the remote sensing image salient target model.

8. The remote sensing image salient target detection method according to claim 7, characterized in that: The specific implementation method of step S4 is: based on the preset indicators PR curve, F value, mean absolute loss and S value, the salient map generated by the remote sensing image salient target model is measured.

Images