Progressive attention augmented optical remote sensing image salient object detection method

Abstract

The application discloses a kind of based on progressive attention enhancement optical remote sensing image salient target detection method, belong to computer vision technical field.The method includes: the original data set is preprocessed;The image after pre-processing is input hierarchical progressive fusion encoder, captures global irregular topological structure and local fine-grained image details, and realizes cross-level feature fusion;The output feature of encoder is input global context enhancement module, adopts parallel multi-branch structure, captures multi-level context information;The output feature of hierarchical progressive fusion encoder and global context enhancement module is input multi-scale progressive attention enhancement decoder, adopts saliency guided attention mechanism, carries out hierarchical decoding to input feature, gradually aggregates deep semantic information and shallow detail features, realizes from coarse to fine progressive optimization, finally generates saliency map.The application can effectively improve the processing performance of irregular topological structure and complex context relationship in optical remote sensing image.

Description

Technical Field

[0001] This invention relates to a salient target detection method for optical remote sensing images based on progressive attention enhancement, belonging to the field of computer vision technology. Background Technology

[0002] Optical Remote Sensing Image Salient Detection (ORSI-SOD) aims to automatically identify the most visually prominent targets or regions in optical remote sensing images (ORSIs) and generate corresponding saliency maps. Unlike detection tasks targeting specific categories, ORSI-SOD is more universal, not limited by specific scenarios, and focuses more on simulating the subjective visual attention mechanism of humans.

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

[0004] First, salient targets in ORSIs often possess irregular topological structures, such as irregular geometries, random spatial distributions, and diverse target scales. For these irregular topologies, global semantic information helps in understanding the overall semantic context, while local detail information plays a crucial role in describing the complex details of salient objects. This necessitates that the encoder simultaneously process global irregular topologies and local fine-grained details, deeply exploring both global and local information to enhance the representation of irregular topologies. Existing convolutional neural network (CNN)-based methods, due to their static network construction and reliance on local feature extraction, lack global context modeling, thus limiting their application in handling various remote sensing scenarios. Transformer-based models can easily establish long-range dependencies, but this leads to the loss of local detail information. Meanwhile, effective fusion of cross-layer features is key to achieving information complementarity and improving model quality. However, existing methods combining Transformer and CNN often neglect the differences in feature importance when fusing multi-scale features, making it difficult to simultaneously optimize spatial detail and semantic consistency, which may lead to noise accumulation.

[0005] Secondly, ORSIs typically possess complex contextual relationships. There are long-range, non-local dependencies between salient targets and their surrounding environment and background elements. Simultaneously, complex component relationships exist within foreground targets. Therefore, fully utilizing the rich contextual information in these images is crucial for improving model quality. However, existing methods have limited ability to model such intricate contextual relationships, ranging from local details to long-range semantics, and are easily affected by complex backgrounds or ignore key long-range relational cues. Common decoders in ORSI-SOD often employ attention mechanisms, lacking dynamic guidance for salient regions. Their single-upsampling predictions or cascaded structures are susceptible to shallow feature noise and lack a saliency-guided progressive optimization mechanism. Summary of the Invention

[0006] To address the aforementioned problems, this invention discloses a salient target detection method for optical remote sensing images based on progressive attention enhancement, specifically comprising the following steps:

[0007] S1: Obtain the dataset for detecting salient targets in optical remote sensing images;

[0008] S2: Preprocessing the salient target detection dataset from optical remote sensing images;

[0009] 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.

[0010] 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;

[0011] 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 its 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.

[0012] 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 a cross-modal attention interaction mechanism.

[0013] S4: Input the output features of the hierarchical progressive fusion encoder into the global context enhancement module, introduce non-local blocks as a pre-module to enhance the spatial relevance of features; adopt a parallel multi-branch structure to capture multi-level contextual information from local details to long-distance semantics, further enhancing the modeling of global contextual 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, and use the saliency-guided attention mechanism to perform hierarchical decoding of the input features;

[0015] 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.

[0016] Each level of decoding unit gradually aggregates deep semantic information and shallow detailed features, achieving progressive optimization from coarse to fine.

[0017] 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.

[0018] Furthermore, the dual-path heterogeneous fusion architecture includes a detail-aware path and a global modeling path; wherein, the detail-aware path is built based on the VGG16 network, removing its terminal global average pooling layer and fully connected layer, and includes five feature extraction stages to capture local detail features of the image; the global modeling path is built based on the Swin Transformer network, including 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.

[0019] Furthermore, the progressive interactive fusion module is specifically embedded in the global modeling path. It performs global average pooling on the input global and local features, upsamples them, and concatenates them with the original features. The concatenated features are then processed by a dimensionality-reducing convolutional layer. Spatial attention weights between global and local features are calculated through cross-modal attention interaction, and complementary features from the detail-aware path and the global modeling path are adaptively fused. Feature reconstruction is performed using the spatial attention map to generate a symmetric attention map for feature selection, further enhancing the semantic consistency of salient targets and thus suppressing noise interference.

[0020] Furthermore, the global context enhancement module specifically comprises: First, for globally irregular topological structures, a nonlocal block is introduced as a pre-module, which establishes long-range associations at any spatial location within the feature map through a self-attention mechanism, thereby enhancing the spatial correlation of features; the nonlocal block utilizes the reduced g, θ, and... The matrix calculates the pixel similarity matrix to generate attention weights, and dynamically weights and fuses global features, thereby breaking the locality limitation of traditional convolution operations and enhancing the consistent representation 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-distance 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, retaining the complementary information of multiple branches, and further enhancing the modeling of global contextual information.

[0021] Furthermore, the parallel multi-branch structure includes: a 1×1 convolution branch, a multi-scale dilated convolution branch, and a global pooling branch. The 1×1 convolution branch performs basic convolution operations while preserving the original spatial information. The multi-scale dilated convolution branch employs three parallel convolutional layers with different dilation rates, which effectively expand the receptive field of the convolution kernel. The three different dilation rates are 3, 6, and 9. The global pooling branch is used to capture image-level semantic information.

[0022] Furthermore, the multi-scale progressive attention enhancement decoder specifically employs a saliency-guided attention mechanism, with each decoding unit containing two parallel multi-scale attention enhancement modules; the multi-scale attention enhancement modules are used to achieve adaptive enhancement and fusion of input features.

[0023] 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, performs bilinear interpolation upsampling on the low-resolution decoded features to achieve cross-level feature alignment; and uses a 3×3 convolutional layer to map the decoded features processed at each level into a single-channel saliency prediction map.

[0024] Output the current decoded features and saliency prediction map to the next level for multi-level supervised training and feature fusion;

[0025] 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.

[0026] Furthermore, the multi-scale attention enhancement module specifically employs a parallel attention enhancement mechanism to compute channel attention and spatial attention in parallel to obtain channel-level and spatial-level attention weights. It then weights and fuses the attention weights of the channel dimension branch and the spatial dimension branch to achieve correlation modeling of the features of the channel dimension branch and the spatial dimension branch, while strengthening the responses of salient regions and boundaries.

[0027] Compared with existing technologies, the present invention has the following advantages:

[0028] 1. This invention proposes a salient target detection method for optical remote sensing images based on progressive attention enhancement, which is used to handle the irregular topological structure and complex contextual relationships of ORSIs.

[0029] 2. This invention designs a progressive interactive encoder, which adopts a dual-path heterogeneous fusion architecture and a hierarchical progressive interactive mechanism to simultaneously capture global irregular topological structures and local fine-grained image details, and achieves cross-level feature fusion.

[0030] 3. This invention constructs a global context enhancement module, introduces non-local blocks as a pre-module to enhance the spatial relevance of features; and adopts a parallel multi-branch structure to capture multi-level contextual information from local details to long-distance semantics, further enhancing the modeling of global contextual information.

[0031] 4. This invention designs a multi-scale progressive attention enhancement decoder, which adopts a saliency-guided attention mechanism to perform cross-level feature fusion and progressive upsampling of the output features of the hierarchical progressive fusion encoder and the global context enhancement module, thereby achieving progressive optimization from coarse to fine. Attached Figure Description

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

[0033] Figure 1 A flowchart illustrating the steps of a salient target detection method for optical remote sensing images based on progressive attention enhancement provided by this invention;

[0034] Figure 2 The network structure diagram of a salient target detection method for optical remote sensing images based on progressive attention enhancement provided by the present invention is shown below.

[0035] Figure 3 This is a structural diagram of the progressive interactive integration module;

[0036] Figure 4 A structural diagram of the global context enhancement module;

[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] In S34, as the network progresses to deeper layers, starting from the third stage (S3) 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-layer feature fusion through a 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. That is: in the first interaction, the features extracted by S3 and V1 are input into the progressive interactive fusion module for fusion, and the fused features are input into S4 and V2 for feature extraction again; in the second interaction, the features extracted by S4 and V2 are input into the progressive interactive fusion module for fusion, and the fused features are input into S5 and V3 for feature extraction again; in the third interaction, the features extracted by S5 and V3 are input into the progressive interactive fusion module for fusion. At this time, the output features of the progressive interactive fusion module are directly input into the global context enhancement module and the multi-scale progressive attention enhancement decoder.

[0048] S35, constructing a progressive interactive fusion module, its schematic diagram is as follows: Figure 3 As shown, this module is embedded in the global modeling path. Through a cross-modal attention interaction mechanism, it adaptively fuses the complementary features of the detail-aware path and the global modeling path to further enhance the semantic consistency of salient targets, thereby suppressing noise interference.

[0049] 1. Global feature map from the Swing Transformer block and local feature maps from VGG16 blocks First, global context enhancement is performed by extracting global context information through global average pooling and concatenating it with the original features. This can be represented as:

[0050]

[0051] in Indicates an upsampling operation. This is a global average pooling operation. [·; ·] indicates feature concatenation, and CBR(·) indicates passing through a conv3×3, BN, and ReLU layer sequentially. This is an element-wise addition.

[0052] 2. Utilize cross-modal attention interaction to compute spatial attention between global and local features:

[0053]

[0054] in and W l i For dimensionality reduction convolutional layers, A i For attention graphs, (·) T For transpose, This refers to matrix multiplication.

[0055] 3. Feature reconstruction using spatial attention maps can be represented as follows:

[0056]

[0057] Where σ(·) represents Softmax normalization, and This indicates that the layers pass through a conv3×3 layer, a BN layer, and a ReLU layer in sequence. This is an element-wise addition.

[0058] 4. Generate symmetric attention maps for feature selection:

[0059]

[0060] in This indicates that the input has passed through a convolutional layer with a kernel size of 3 and a stride of 1, and σ(·) represents softmax normalization. This is an element-wise addition.

[0061] 5. The final output is the fused feature. Attention map Among them, fusion features and They are fed into the Swing Transformer block and the VGG16 block, respectively.

[0062] S4: The feature map output from the hierarchical progressive fusion encoder. Input the global context enhancement module.

[0063] like Figure 4 The diagram shown illustrates the structure of a global context enhancement module according to an embodiment of the present invention. Step S4 specifically includes steps S41-S45:

[0064] S41, for globally irregular topologies, a nonlocal block is introduced as a pre-module. A self-attention mechanism is used to establish long-range associations at any spatial location within the feature map, enhancing the spatial correlation of features. The nonlocal block utilizes the reduced g, θ, and... The matrix calculation of pixel similarity matrix generates attention weights, and dynamically weights and fuses global features, thereby breaking the locality limitation of traditional convolution operations and enhancing the consistent representation of features in occluded regions and scattered targets; it can be described as follows:

[0065]

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

[0067] S42 uses a 1×1 convolution branch to perform basic convolution operations, preserving the original spatial information:

[0068] Y1=ReLU(BN(C 1×1 (Y0)))

[0069] S43, the multi-scale dilated convolution branch employs three parallel convolutional layers with different dilation rates. These different dilation rates effectively expand the receptive field of the convolutional kernel; the three different dilation rates are 3, 6, and 9.

[0070]

[0071] S44 uses global pooling branches to capture image-level semantic information:

[0072]

[0073] in Indicates upsampling, This is global average pooling.

[0074] S45. Finally, feature concatenation and compression are performed. The output features of each branch are concatenated along the channel dimension, and enhanced features X are generated through a feature fusion layer. This not only preserves the complementary information of multiple branches but also reduces computational redundancy through dimensionality reduction, ensuring the efficiency of subsequent decoding stages.

[0075]

[0076] Where || represents feature concatenation along the channel dimension, Droput 0.5 This is a dimensionality 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 enhancement decoder.

[0078] like Figure 5 The diagram shows a multi-scale progressive attention enhancement decoder structure according to an embodiment of the present invention. Step S5 specifically includes steps S51-S59:

[0079] S51, the multi-scale progressive attention-enhanced decoder, employs a saliency-guided attention mechanism to perform hierarchical decoding of input features; it includes five levels, with each decoding unit containing two parallel multi-scale attention enhancement modules;

[0080] S52, the multi-scale attention enhancement module employs a parallel attention enhancement mechanism, focusing on the backbone features of the current level. (or jump characteristics) (and the previous level significance prediction plot SA) i+1 Parallel computation of channel attention and spatial attention is used to obtain spatial and channel-level attention weights; the correlation modeling of features in spatial and channel dimensions is realized, while enhancing the responses of salient regions and boundaries;

[0081] S53, in the channel dimension branch, first, the saliency prediction plot... Flattening is performed to obtain the significant response factor for each channel. Main characteristics Create a global max pool and flatten it out. Then multiply with the significance factor matrix to highlight channels with high significance responses:

[0082]

[0083]

[0084] in For flattening operation, For global max pooling; then, obtain M through MLP. CA Then, the attention weights are applied to the output channel via the Sigmoid algorithm:

[0085]

[0086] in For matrix multiplication, ∮(·) is the Sigmoid function; finally, M is then... c With main characteristics Residual connections ensure stable training and preserve original information.

[0087]

[0088] in ⊙ represents element-wise addition, and ⊙ represents element-wise multiplication.

[0089] S54, in the spatial dimension branch, firstly, spatial max pooling is used to process the backbone features. Channel compression is performed to obtain a compressed image:

[0090]

[0091] Where SMP(·) is spatial max pooling; subsequently, it is combined with the saliency prediction map SA i+1 Channel-dimensional concatenation is performed, followed by a 3×3 convolution and a sigmoid function for convolution projection, resulting in the spatial attention map M. s :

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

[0093] Where ∮(·) is the Sigmoid function, and Concat(·) is the feature concatenation; then, feature reweighting is performed to highlight the target location, and the formula is as follows:

[0094]

[0095] in ⊙ represents element-wise addition, and ⊙ represents 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, each decoding unit concatenates the outputs of the multi-scale attention enhancement modules at levels i and i+1. Then, cross-level feature alignment is performed, followed by two 3×3 convolutions to complete information fusion and channel number adjustment. The low-resolution decoded features are then upsampled proportionally to the same spatial resolution as the corresponding layer of the backbone using bilinear interpolation to obtain the current decoded features. Provide matching feature sizes for subsequent fusion:

[0100]

[0101] in For upsampling operations, BConv(·) performs conv3×3, BN, and ReLU operations, and Concat(·) performs feature concatenation.

[0102] S57, after another 3×3 convolution, maps the multi-channel decoded features to a single-channel saliency prediction map SA. i :

[0103]

[0104] Where ∮(·) is the Sigmoid function;

[0105] S58 outputs the current decoded features and saliency prediction map to the next level for multi-level supervised training and feature fusion;

[0106] S59, each decoding unit generates supervisory signals from coarse to fine step by step, and performs upsampling step by step to achieve multi-scale feature aggregation and attention-guided enhancement.

[0107] 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.

[0108] Table 1 compares the proposed method with the well-known ORSI-SOD method on the ORSSD, EORSSD, and ORSI-4199 datasets (S in the table...). α F β E ε A higher MAE value indicates better model performance, while a lower MAE value indicates better model performance.

[0109] Table 1. Comparative experimental results on different datasets

[0110]

[0111] As shown in Table 1, this embodiment outperforms existing methods in multiple metrics across the three datasets, demonstrating the effectiveness of the method presented in this embodiment.

[0112] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various modifications or variations within the scope of the claims, which do not affect the essence of the present invention. The above preferred features can be used in any combination without conflict.

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, specifically: 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. 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 best 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 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.

3. 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.

4. 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.

5. 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.

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 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