Optical remote sensing image salient target detection method based on progressive attention enhancement

Abstract

The invention discloses an optical remote sensing image salient target detection method based on progressive attention enhancement, and belongs to the technical field of computer vision. The method comprises the following steps: preprocessing an original data set; inputting the preprocessed image into a hierarchical progressive fusion encoder, capturing a global irregular topological structure and local fine-grained image details, and realizing cross-hierarchical feature fusion; inputting the output characteristics of the encoder into a global context enhancement module, and capturing multi-level context information by adopting a parallel multi-branch structure; and inputting the output features of the hierarchical progressive fusion encoder and the global context enhancement module into a multi-scale progressive attention enhancement decoder, carrying out hierarchical decoding on the input features by adopting a saliency-guided attention mechanism, and gradually aggregating deep semantic information and shallow detail features to realize coarse-to-fine progressive optimization, so as to improve the robustness of the multi-scale progressive attention enhancement decoder. And finally generating a saliency map. The method can effectively improve the processing performance of an irregular topological structure and a complex context relationship in the optical remote sensing image.

Description

TECHNICAL FIELD

[0001] The present application relates to a kind of optical remote sensing image salient target detection method based on progressive attention enhancement, belong to computer vision technical field. BACKGROUND

[0002] Optical remote sensing image salient target detection (ORSI-SOD) aims to automatically identify the most visually prominent target or region in optical remote sensing image (ORSIs), and generate the corresponding saliency map. Unlike detection tasks for specific categories, ORSI-SOD is more universal and is not limited by specific scenes, and is more focused on simulating human subjective visual attention mechanisms.

[0003] Although current ORSI-SOD methods have made some progress in improving detection performance, they are generally difficult to effectively deal with the inherent challenges of ORSIs.

[0004] First, the salient targets in ORSIs usually have irregular topological structures, such as irregular geometric shapes, random spatial distribution, and diversified target scales. For these irregular topological structures, global semantic information helps to understand the overall semantic context, while local detail information plays a key role in describing the complex details of salient objects, which requires the encoder to be able to handle both global irregular topological structures and local fine-grained details, and to explore global and local information to enhance the representation of irregular topologies. Existing convolutional neural network (CNN) based methods lack global context modeling due to their static network construction and reliance on local feature extraction, which limits their ability to handle multiple scenarios in remote sensing. Transformer-based models can easily establish long-range dependencies, but can lead to the loss of local detail information. At the same time, effective fusion of cross-layer features is key to achieving information complementarity and improving model quality. However, existing methods that combine Transformers and CNNs often ignore the importance differences of features when fusing multi-scale features, making it difficult to optimize both spatial details and semantic consistency, which can lead to noise accumulation.

[0005] Secondly, ORSIs usually have complex contextual relationships. There are long-range, non-local dependencies between the salient object and its surrounding environment as well as background elements. Meanwhile, there are also complex component associations within the foreground object. Therefore, it is essential to make full use of the rich contextual information in these images to improve the quality of the model. However, existing methods have limited ability to model such complex contextual relationships from local details to long-range semantics, and are easily disturbed by complex backgrounds or ignore key long-range association clues. The common decoder in ORSI-SOD usually adopts an attention mechanism, lacks dynamic guidance for salient regions, and its single upsampling prediction or cascading structure is easily disturbed by shallow feature noise and lacks a progressive optimization mechanism for saliency guidance. SUMMARY

[0006] To solve the above problems, the application discloses an optical remote sensing image salient object detection method based on progressive attention enhancement, which specifically comprises the following steps:

[0007] S1: obtaining an optical remote sensing image salient object detection dataset;

[0008] S2: preprocessing the optical remote sensing image salient object detection dataset;

[0009] S3: inputting the preprocessed image into a hierarchical progressive fusion encoder to extract features, capture global irregular topological structures and local fine-grained image details; at the same time, using the progressive interactive fusion module in the hierarchical progressive fusion encoder, bridging the semantic differences between multi-scale features from the global and local, realizing cross-level feature fusion;

[0010] The hierarchical progressive fusion encoder adopts a dual-path heterogeneous fusion architecture and a hierarchical progressive fusion mechanism to realize feature extraction and fusion;

[0011] The hierarchical progressive fusion mechanism is that the preprocessed image is input into the hierarchical progressive fusion encoder, first passes through the feature extraction of the detail perception path, and as the network develops to a deeper layer, from the third stage of the detail perception path, starts to process the input features of the same layer in a parallel manner with the three stages of the global modeling path, the cross-level feature fusion between the two paths is realized through the progressive interactive fusion module, the local fine-grained image details of the detail perception path and the global irregular topological structure of the global modeling path are bidirectionally compensated layer by layer, and high-quality features are gradually purified and the semantic gap is bridged;

[0012] The progressive interactive fusion module is embedded in the global modeling path, and through the cross-modal attention interaction mechanism, the complementary features of the detail perception path and the global modeling path are adaptively fused;

[0013] S4: input the output features of the hierarchical progressive fusion encoder into the global context enhancement module, introduce a non-local block as a pre-module to enhance the spatial correlation of the features; adopt a parallel multi-branch structure to capture multi-level context information from local details to long-distance semantics and further enhance the modeling of global context information;

[0014] S5: input the output features of the hierarchical progressive fusion encoder and the global context enhancement module into the multi-scale progressive attention enhancement decoder, adopt a saliency-guided attention mechanism to hierarchically decode the input features;

[0015] Each level of the decoding unit uses a multi-scale attention enhancement module as a core fusion unit, and the multi-scale attention enhancement module adopts a parallel attention enhancement mechanism to simultaneously model the spatial and channel dimension feature correlation, strengthen the salient region and boundary response, and realize adaptive enhancement and fusion of the input features;

[0016] Each level of the decoding unit gradually aggregates deep semantic information and shallow detail features to realize progressive optimization from coarse to fine;

[0017] S6: train the model on the training set and save the best model parameters; input the test set image into the model to obtain the salient target detection result.

[0018] Further, the dual-path heterogeneous fusion architecture includes a detail perception path and a global modeling path; wherein the detail perception path is constructed based on a VGG16 network, removes the end global average pooling layer and the fully connected layer, includes five feature extraction stages, and is used to capture local detail features of an image; the global modeling path is constructed based on a Swin Transformer network, includes three feature extraction stages, integrates a progressive interactive fusion module, is used to dynamically model the global irregular topological structure of an image, and realizes cross-level feature fusion.

[0019] Further, the progressive interactive fusion module is specifically: the progressive interactive fusion module is embedded in the global modeling path, performs a global average pooling operation on the input global features and local features, and after upsampling, the original features are spliced; the spliced features are processed by a dimension reduction convolution layer, the spatial attention weights between the global and local features are calculated through cross-modal attention interaction, the complementary features of the detail perception path and the global modeling path are adaptively fused; the spatial attention map is used for feature reconstruction to generate a symmetric attention map for feature selection, further enhancing the semantic consistency of the salient target, thereby suppressing noise interference.

[0020] Further, the global context enhancement module, specifically: firstly, for the global irregular topology, a non-local block is introduced as a pre-module, and long-range correlation between any spatial positions in the feature map is established through self-attention mechanism to enhance the spatial correlation of the features; the non-local block uses g, θ and The matrix calculates the similarity matrix between pixels, generates attention weights, and dynamically weights and fuses global features, so as to break the locality restriction of traditional convolution operation, enhance the consistent expression of features of the occluded area and scattered target; then, a parallel multi-branch structure is adopted to capture multi-level context information from local details to remote semantics; finally, feature splicing and compression are performed, the output features of each branch are spliced along the channel dimension, and enhanced features are generated through a feature fusion layer to retain the complementary information of the multi-branch and further enhance the modeling of the global context information.

[0021] Further, the parallel multi-branch structure includes a 1x1 convolution branch, a multi-scale dilated convolution branch and a global pooling branch, wherein the 1x1 convolution branch performs basic convolution operation to retain original spatial information; the multi-scale dilated convolution branch adopts three parallel convolution layers with different expansion rates, and the dilated convolution with different expansion rates effectively expands the receptive field of the convolution kernel; wherein the three different expansion rates are 3, 6 and 9; the global pooling branch is used to capture image-level semantic information.

[0022] Further, the multi-scale progressive attention enhancement decoder, specifically: an attention mechanism guided by saliency is adopted, and each decoding unit includes two parallel multi-scale attention enhancement modules; the multi-scale attention enhancement module is used to realize adaptive enhancement and fusion of input features;

[0023] The output of the two multi-scale attention enhancement modules in each decoding unit is spliced; two serial 3x3 convolution layers are used for information fusion and channel adjustment, bilinear interpolation up-sampling is performed on the low-resolution decoding features to realize cross-level feature alignment; a 3x3 convolution layer is used to map the decoding features processed at each level to a single-channel saliency prediction map;

[0024] The current decoding feature and the saliency prediction map are output to the next level for multi-level supervised training and feature fusion;

[0025] Each decoding unit generates a multi-scale saliency prediction map from coarse to fine, and gradually aggregates deep semantic information and shallow detail features through step-by-step up-sampling, realizing multi-scale feature fusion and attention guided enhancement.

[0026] Further, the multi-scale attention enhancement module, specifically: adopt parallel attention enhancement mechanism, parallel computing channel attention and spatial attention, to obtain channel level and spatial level attention weight, and the attention weight of channel dimension branch and spatial dimension branch is weighted and fused, the correlation modeling of channel dimension branch and spatial dimension branch feature is realized, and the significant area and boundary response are strengthened.

[0027] Compared with the prior art, the present application has the following beneficial results:

[0028] 1. The present application provides a kind of optical remote sensing image salient target detection method based on progressive attention enhancement, for processing ORSIs irregular topological structure and complex context relationship.

[0029] 2. The present application designs progressive interactive encoder, adopts double-path heterogeneous fusion architecture and hierarchical progressive interaction mechanism, simultaneously captures global irregular topological structure and local fine-grained image details, and realizes cross-level feature fusion.

[0030] 3. The present application constructs global context enhancement module, introduces non-local block as front-end module, enhances the spatial correlation of feature;Adopt parallel multi-branch structure, capture multi-level context information from local details to long-distance semantics, further enhance the modeling of global context information.

[0031] 4. The present application designs multi-scale progressive attention enhancement decoder, adopts saliency guided attention mechanism, and the output features of hierarchical progressive fusion encoder and global context enhancement module are cross-level feature fusion and step-by-step up-sampling, realize the progressive optimization from coarse to fine. BRIEF DESCRIPTION OF DRAWINGS

[0032] Other features, objects and advantages of the present application will become more apparent from the following detailed description of non-limiting embodiments, read in conjunction with the accompanying drawings:

[0033] Figure 1 A step flow chart of the optical remote sensing image salient target detection method based on progressive attention enhancement provided by the present application is provided;

[0034] Figure 2 A network structure diagram of the optical remote sensing image salient target detection method based on progressive attention enhancement provided by the present application is provided;

[0035] Figure 3 A structure diagram of progressive interactive fusion module is provided;

[0036] Figure 4 A structure diagram of global context enhancement module is provided;

[0037] Figure 5This is a structural diagram of a multi-scale progressive attention enhancement decoder. Detailed Implementation

[0038] To enable those skilled in the art to better understand the present invention, the technical solution of the present invention will be described in detail below with reference to specific embodiments. The specific embodiments described herein are only for explaining the present invention and are not intended to limit the present invention.

[0039] like Figure 1 The diagram shows a flowchart of a salient target detection method for optical remote sensing images based on progressive attention enhancement, according to an embodiment of the present invention, including steps S1-S6:

[0040] S1: Obtain the optical remote sensing image salient target detection datasets ORSSD, EORSSD, and ORSI-4199;

[0041] S2: Preprocessing the salient target detection dataset from optical remote sensing images;

[0042] S3: Input the preprocessed image into the hierarchical progressive fusion encoder for feature extraction, capturing global irregular topological structures and local fine-grained image details; at the same time, the progressive interactive fusion module in the hierarchical progressive fusion encoder is used to bridge the semantic differences between multi-scale features from global and local sources, and achieve cross-level feature fusion.

[0043] like Figure 2 The diagram shown illustrates the network architecture of a salient target detection method for optical remote sensing images based on progressive attention enhancement, according to an embodiment of the present invention. The method includes a hierarchical progressive fusion encoder, a global context enhancement module, and a multi-scale progressive attention enhancement decoder. The backbone network is constructed based on VGG16 and Swin Transformer, and is divided into a detail-aware path and a global modeling path. Step S3 specifically includes steps S31-S35:

[0044] S31, the preprocessed image is input into the hierarchical progressive fusion encoder, and features are extracted first through the detail-aware path;

[0045] S32, Detail-Aware Path: Based on the VGG16 network, its terminal global average pooling layer and fully connected layer are removed. It contains five feature extraction stages to capture local detail features of the image.

[0046] S33, Global Modeling Path: Based on the Swing Transformer network, it includes three feature extraction stages and integrates a progressive interactive fusion module to dynamically model the global irregular topology of the image and achieve cross-level feature fusion.

[0047] S34, with the network developing to a deeper level, from the third stage (S3) of the detail-aware path, the input features of the same layer are processed in a parallel manner with the three stages of the global modeling path, and the cross-level feature fusion between the two paths is realized through the progressive interaction fusion module, the local fine-grained image details of the detail-aware path and the global irregular topological structure of the global modeling path are bidirectionally compensated layer by layer, that is: the first interaction, the features extracted by S3 and V1 are input into the progressive interaction fusion module for fusion, and the fused features are input into S4 and V2 respectively for feature extraction again; the second interaction, the features extracted by S4 and V2 are input into the progressive interaction fusion module for fusion, and the fused features are input into S5 and V3 respectively for feature extraction again; the third interaction, the features extracted by S5 and V3 are input into the progressive interaction fusion module for fusion, and the output features of the progressive interaction fusion module are directly input into the global context enhancement module and the multi-scale progressive attention enhancement decoder respectively.

[0048] S35, the progressive interaction fusion module is constructed, and its schematic diagram is shown in Figure 3 The module is embedded in the global modeling path, and the complementary features of the detail-aware path and the global modeling path are adaptively fused through the cross-modal attention interaction mechanism, so as to further enhance the semantic consistency of the salient target and inhibit noise interference:

[0049] 1. Global feature map from Swin Transformer block and local feature map from VGG16 block First, global context enhancement is performed, global context information is extracted through global average pooling, and is spliced with the original feature, which can be expressed as:

[0050]

[0051] Wherein represents the up-sampling operation, is the global average pooling operation, [·; ·] represents the feature splicing, CBR(·) represents the conv3x3, BN and ReLU layers in turn, is the element-wise addition.

[0052] 2. The spatial attention between global and local features is calculated by cross-modal attention interaction:

[0053]

[0054] Wherein and W l i are dimension reduction convolution layers, A i is an attention map, (·) T is transposed, For matrix multiplication.

[0055] 3. Feature reconstruction is performed by using the spatial attention map, and the overall process can be represented as follows:

[0056]

[0057] wherein sigma (·) represents Softmax normalization, and represents sequentially passing through a conv3x3, BN and ReLU layer, is an element-wise addition.

[0058] 4. A symmetric attention map is generated for feature selection:

[0059]

[0060] wherein represents passing through a convolutional layer with a kernel size of 3 and a step size of 1, sigma (·) represents softmax normalization, is an element-wise addition.

[0061] 5. The final output is a fusion feature and an attention map wherein the fusion feature and are respectively fed into a Swin Transformer block and a VGG16 block.

[0062] S4: The feature map output by the hierarchical progressive fusion encoder is input into a global context enhancement module.

[0063] As shown in Figure 4 , a global context enhancement module structure diagram of an embodiment provided by the application is shown, and the specific steps of step S4 include S41-S45:

[0064] S41, for a global irregular topological structure, a non-local block is introduced as a front-end module, long-range association between any spatial positions in the feature map is established through a self-attention mechanism, and the spatial correlation of the features is enhanced; the non-local block uses g, theta and matrix to calculate the similarity matrix between pixels, generates attention weights, and dynamically weights and fuses global features, thereby breaking the locality limitation of traditional convolutional operations and enhancing the consistent expression of features of the occluded area and the scattered target; which can be expressed as follows:

[0065]

[0066] wherein F is the feature map output by the hierarchical progressive fusion encoder For matrix multiplication, For element-wise addition.

[0067] S42, the basic convolution operation is performed by using a 1*1 convolution branch, and the original spatial information is reserved:

[0068] Y1 = ReLU(BN(C 1×1 (Y0))

[0069] S43, the multi-scale dilated convolution branch adopts three parallel convolution layers with different expansion rates, and the receptive field of the convolution kernel is effectively expanded by the dilated convolution with different expansion rates; wherein the three different expansion rates are 3, 6 and 9:

[0070]

[0071] S44, a global pooling branch is used to capture image-level semantic information:

[0072]

[0073] wherein represents up-sampling, is global average pooling.

[0074] S45, finally, feature splicing and compression are performed, the output features of each branch are spliced along the channel dimension, and enhanced features X are generated through a feature fusion layer; not only the complementary information of the multi-branch is reserved, but also the calculation redundancy is reduced through dimension reduction to ensure the efficiency of the subsequent decoding stage:

[0075]

[0076] wherein || represents feature splicing in the channel dimension, Droput 0.5 is a dimension reduction operation.

[0077] S5: input the output features of the hierarchical progressive fusion encoder and the enhanced features of the global context enhancement module into the multi-scale progressive attention enhanced decoder.

[0078] As Figure 5 shown, the multi-scale progressive attention enhanced decoder structure diagram of an embodiment provided by the application, the specific steps of step S5 include S51-S59:

[0079] S51, the multi-scale progressive attention enhanced decoder adopts a saliency-guided attention mechanism to perform hierarchical decoding on the input features; a total of five levels, and each level decoding unit includes two parallel multi-scale attention enhanced modules;

[0080] S52, the multi-scale attention enhanced module adopts a parallel attention enhanced mechanism, and respectively processes the main features (or jump feature ) and the previous level of saliency prediction map SA i+1 Parallelly compute channel attention and spatial attention to obtain attention weights at spatial level and channel level; realize the correlation modeling of spatial dimension branch and channel dimension branch features, and strengthen the salient region and boundary response at the same time;

[0081] S53, in the channel dimension branch, first flatten the saliency prediction map SA to obtain the salient response factor of each channel Global max-pooling is performed on the backbone feature and flattened into , which is then multiplied by the saliency factor matrix to highlight the channels with high saliency response:

[0082]

[0083]

[0084] wherein is the flattening operation, is the global max-pooling; then, M CA is obtained through MLP, and the channel attention weight is output through Sigmoid:

[0085]

[0086] wherein is matrix multiplication, and ∮(·) is the Sigmoid function; finally, M c is combined with the backbone feature through residual connection to stabilize the training and retain the original information:

[0087]

[0088] wherein is element-wise addition, and ⊙ is element-wise multiplication.

[0089] S54, in the spatial dimension branch, first compress the backbone feature through spatial max-pooling to obtain a compressed map:

[0090]

[0091] wherein SMP(·) is spatial max-pooling; subsequently, it is concatenated with the saliency prediction map SA i+1 , and then sent into a layer of 3×3 convolution and Sigmoid function for convolution projection to obtain the spatial attention map M s :

[0092] M s =∮(Conv 3×3 (Concat(X s ,SA i+1 ))

[0093] where ∮(·) is a Sigmoid function, and Concat(·) is feature concatenation; then, feature re-weighting is performed to highlight the target position, and the formula is as follows:

[0094]

[0095] where is element-wise addition, and is element-wise multiplication;

[0096] S55, the attention weights of the channel dimension branch and the spatial dimension branch are weighted and fused to obtain the fused feature The formula is as follows:

[0097]

[0098] where α is set to 0.7, and β is set to 0.3;

[0099] S56, the i-th multi-scale attention enhancement module and the i+1-th multi-scale attention enhancement module are feature concatenated; then, cross-level feature alignment is performed, and information fusion and channel number adjustment are completed through two layers of 3x3 convolution; the low-resolution decoding feature is upsampled to the same spatial resolution as the corresponding layer of the backbone in a bilinear interpolation manner, to obtain the current decoding feature to provide matching feature size for subsequent fusion:

[0100]

[0101] where is an upsampling operation, BConv(·) is a conv3x3, BN and ReLU operation, and Concat(·) is feature concatenation;

[0102] S57, a layer of 3x3 convolution is further performed, and the multi-channel decoding feature is mapped to a single-channel saliency prediction map SA i :

[0103]

[0104] where ∮(·) is a Sigmoid function;

[0105] S58, the current decoding feature and the saliency prediction map are output to the next level for multi-level supervised training and feature fusion;

[0106] S59, each level of the decoding unit forms a supervision signal from coarse to fine, and performs upsampling step by step, realizing multi-scale feature aggregation and attention guiding enhancement.

[0107] S6: training the model on the training set, saving the best model parameters; inputting the test set image into the model to obtain the significant target detection result.

[0108] Table 1 is a comparison of the proposed method with the well-known ORSI-SOD method on ORSSD, EORSSD, ORSI-4199 data sets (S α , F β , E ε The higher the value, the better the model performance, and the lower the MAE value, the better the model performance

[0109] Table 1 shows the results of comparative experiments on different data sets

[0110]

[0111] As shown in Table 1, the proposed method outperforms the existing method in multiple indicators on three data sets, proving the effectiveness of the proposed method.

[0112] The specific embodiments of the present application are described above. It should be understood that the present application is not limited to the above specific embodiments, and those skilled in the art can make various modifications or modifications within the scope of the claims, which does not affect the essential content of the present application. The above preferred features can be used in combination in the case of not conflicting with each other.

Claims

1. A salient target detection method for optical remote sensing images based on progressive attention enhancement, characterized in that, Includes the following steps: S1: Obtain the dataset for detecting salient targets in optical remote sensing images; S2: Preprocessing the salient target detection dataset from optical remote sensing images; S3: Input the preprocessed image into the hierarchical progressive fusion encoder for feature extraction, capturing global irregular topological structures and local fine-grained image details; at the same time, the progressive interactive fusion module in the hierarchical progressive fusion encoder is used to bridge the semantic differences between multi-scale features from global and local sources, and achieve cross-level feature fusion. The hierarchical progressive fusion encoder described above uses a dual-path heterogeneous fusion architecture and a hierarchical progressive fusion mechanism to achieve feature extraction and fusion; The aforementioned dual-path heterogeneous fusion architecture is divided into a detail-aware path and a global modeling path; The hierarchical progressive fusion mechanism is as follows: the preprocessed image is input into the hierarchical progressive fusion encoder. First, it is processed for feature extraction through the detail-aware path. As the network develops to a deeper level, starting from the third stage of the detail-aware path, it begins to process the input features of the same layer in parallel with the three stages of the global modeling path. The two paths achieve cross-level feature fusion through the progressive interactive fusion module. The local fine-grained image details of the detail-aware path and the global irregular topology of the global modeling path are compensated bidirectionally layer by layer, gradually purifying high-quality features and bridging the semantic gap. The progressive interaction fusion module is embedded in the global modeling path, and adaptively fuses the complementary features of the detail-aware path and the global modeling path through cross-modal attention interaction. S4: Input the output features of the hierarchical progressive fusion encoder into the global context enhancement module, and introduce non-local blocks as a pre-module to enhance the spatial correlation of features; A parallel multi-branch structure is adopted to capture multi-level contextual information from local details to long-range semantics, further enhancing the modeling of global contextual information; S5: Input the output features of the hierarchical progressive fusion encoder and the global context enhancement module into the multi-scale progressive attention enhancement decoder, and use the saliency-guided attention mechanism to perform hierarchical decoding of the input features; Each decoding unit utilizes a multi-scale attention enhancement module as the core fusion unit. The multi-scale attention enhancement module adopts a parallel attention enhancement mechanism, while simultaneously modeling the feature correlation between spatial and channel dimensions, strengthening the responses of salient regions and boundaries, and realizing adaptive enhancement and fusion of input features. Each level of decoding unit gradually aggregates deep semantic information and shallow detailed features, achieving progressive optimization from coarse to fine. S6: Train the model on the training set and save the optimal model parameters; input the test set images into the model to obtain the salient object detection results.

2. The salient target detection method for optical remote sensing images based on progressive attention enhancement according to claim 1, characterized in that, The dual-path heterogeneous fusion architecture includes: Detail-aware path: Based on the VGG16 network, its terminal global average pooling layer and fully connected layer are removed, and it contains five feature extraction stages to capture local detail features of the image; Global modeling path: Based on the Swing Transformer network, it includes three feature extraction stages and integrates a progressive interactive fusion module to dynamically model the global irregular topology of the image and achieve cross-level feature fusion.

3. The salient target detection method for optical remote sensing images based on progressive attention enhancement according to claim 1, characterized in that, The progressive interactive fusion module is constructed as follows: The progressive interactive fusion module is embedded in the global modeling path, which performs global average pooling on the input global and local features respectively, and then concatenates them with the original features after upsampling. The concatenated features are processed by a dimensionality-reducing convolutional layer. Spatial attention weights between global and local features are calculated through cross-modal attention interaction. Complementary features from detail-aware paths and global modeling paths are adaptively fused. The spatial attention map is used for feature reconstruction to generate a symmetric attention map for feature selection, which further enhances the semantic consistency of salient targets and thus suppresses noise interference.

4. The salient target detection method for optical remote sensing images based on progressive attention enhancement according to claim 1, characterized in that, The global context enhancement module is constructed as follows: First, for globally irregular topological structures, nonlocal blocks are introduced as pre-modules, and long-range associations at any spatial location within the feature map are established through a self-attention mechanism to enhance the spatial correlation of features. Nonlocal blocks utilize channel-reduced g, θ, and The matrix calculates the similarity matrix between pixels, generates attention weights, and dynamically weights and fuses global features, thereby breaking the locality limitation of traditional convolution operations and enhancing the consistent expression of features for occluded regions and scattered targets. Subsequently, a parallel multi-branch structure is adopted to capture multi-level contextual information from local details to long-range semantics. Finally, feature concatenation and compression are performed, concatenating the output features of each branch along the channel dimension, generating enhanced features through a feature fusion layer, preserving the complementary information of multiple branches, and further enhancing the modeling of global contextual information.

5. The salient target detection method for optical remote sensing images based on progressive attention enhancement according to claim 1, characterized in that, The parallel multi-branch structure includes: 1×1 convolution branch: Performs basic convolution operations while preserving the original spatial information; Multi-scale dilated convolution branch: Three parallel convolutional layers with different dilation rates are used. The dilated convolution with different dilation rates effectively expands the receptive field of the convolutional kernel; the three different dilation rates are 3, 6 and 9 respectively. Global pooling branch: used to capture image-level semantic information.

6. The salient target detection method for optical remote sensing images based on progressive attention enhancement as described in claim 1, characterized in that, The multi-scale progressive attention-enhanced decoder is constructed as follows: A saliency-guided attention mechanism is adopted, and each decoding unit contains two parallel multi-scale attention enhancement modules; the multi-scale attention enhancement modules are used to achieve adaptive enhancement and fusion of input features; Each decoding unit concatenates the outputs of the two multi-scale attention enhancement modules; it uses two cascaded 3×3 convolutional layers for information fusion and channel adjustment, and performs bilinear interpolation upsampling on the low-resolution decoded features to achieve cross-level feature alignment. A 3×3 convolutional layer is used to map the decoded features after each level of processing into a single-channel saliency prediction map; Output the current decoded features and saliency prediction map to the next level for multi-level supervised training and feature fusion; Each decoding unit generates multi-scale saliency prediction maps from coarse to fine, and gradually aggregates deep semantic information and shallow detail features through progressive upsampling, thereby achieving multi-scale feature fusion and attention-guided enhancement.

7. The salient target detection method for optical remote sensing images based on progressive attention enhancement as described in claim 1, characterized in that, The multi-scale attention enhancement module is constructed as follows: A parallel attention enhancement mechanism is adopted to compute channel attention and spatial attention in parallel to obtain channel-level and spatial-level attention weights. The attention weights of the channel dimension branch and the spatial dimension branch are then weighted and fused to achieve the correlation modeling of the features of the channel dimension branch and the spatial dimension branch, while strengthening the response of salient regions and boundaries.

Images