Remote sensing image directed target detection method based on double-domain feature fusion

Abstract

The invention relates to the technical field of remote sensing target detection, in particular to a remote sensing image directed target detection method based on double-domain feature fusion, and the method comprises the steps: obtaining a remote sensing image data set, and obtaining a training set; constructing a remote sensing target detection network, wherein the remote sensing target detection network comprises a feature extraction branch, a double-domain feature fusion branch for performing feature fusion on spatial domain features and frequency domain features of the remote sensing images, and a target detection branch for judging corresponding categories and positions of the remote sensing images according to different fusion features; training a remote sensing target detection network by using the training set to obtain a remote sensing target detection model; and inputting a to-be-detected remote sensing image into the remote sensing target detection model, and outputting a corresponding detection category. According to the method, the effect of the spatial domain and the frequency domain on remote sensing classification is fully considered, and through adaptive selection of the spatial domain and the frequency domain and feature interaction fusion between the two domains, fusion of global context information and local information is effectively enhanced, and the deficiency of target information is made up.

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing target detection, and in particular relates to a remote sensing image directional target detection method based on dual-domain feature fusion. Background Art

[0002] At present, general target detection algorithms in natural scenes have achieved excellent performance, and their robustness and generalization capabilities have been well verified. However, there are certain differences in the target characteristics between remote sensing images and natural scene images, such as remote sensing images have a wide field of view, different target directions, uneven distribution, and weak feature expression.

[0003] Therefore, a large number of researchers have made improvements and designs based on the transfer algorithm for different characteristics of remote sensing images. High-performance remote sensing target detectors usually rely on the RCNN framework, which consists of a region proposal network and a detection head. RPN proposes high-quality regions of interest from the backbone feature map, while the regional neural network detection head is responsible for object classification and bounding box regression. In recent years, several variants of the RCNN (Region-based convolutional neural networks) framework have been proposed. Such methods have improved the performance of directed target detection in remote sensing images by improving and designing various modules based on convolutional neural networks. However, feature extraction through a single spatial domain limits the expression of target features and cannot exert greater potential in remote sensing target detection tasks.

[0004] In order to solve the problem of complex and changeable background and weak target feature expression in remote sensing images, the current method of extracting target features only through spatial domain is difficult to fully utilize the feature information, resulting in poor target detection and positioning accuracy. Summary of the invention

[0005] In view of this, the present invention aims to provide a remote sensing image directional target detection method based on dual-domain feature fusion, which fully utilizes the complementary features between different domains through adaptive selection of spatial domain and frequency domain and interactive fusion of features between the dual domains, effectively enhances the fusion of global context information and local information, and makes up for the lack of target information.

[0006] To achieve the above object, the technical solution created by the present invention is implemented as follows: A method for detecting oriented targets in remote sensing images based on dual-domain feature fusion, comprising: S1: Obtain a remote sensing image dataset and preprocess the remote sensing image dataset to obtain a training set; S2: Construct a remote sensing target detection network; wherein the remote sensing target detection network includes: The feature extraction branch extracts initial features of different sizes from the input remote sensing image; The dual-domain feature fusion branch extracts the corresponding spatial domain features and frequency domain features from the initial features of different sizes; the spatial domain features and frequency domain features of the corresponding sizes are fused to obtain the corresponding fused features; The target detection branch determines the category and location of the remote sensing image based on different fusion features; S3: Using the training set obtained in step S1 to train the remote sensing target detection network in step S2, to obtain a remote sensing target detection model; S4: Input the remote sensing image to be detected into the remote sensing target detection model of step S3, and output the corresponding detection category and position.

[0007] Furthermore, the dual-domain feature fusion branch in step S2 includes: a spatial domain adaptive selection module, which extracts corresponding spatial domain features from the initial features; a frequency domain adaptive selection module, which extracts corresponding frequency domain features from the initial features; a dual-domain feature interaction module, which fuses spatial domain features and frequency domain features of corresponding sizes to obtain corresponding preliminary fused features; and a feature pyramid module, which extracts multi-scale information from the preliminary fused features to obtain fused features.

[0008] Furthermore, the spatial domain adaptive selection module includes a feature extraction submodule, a pooling submodule, a multi-scale extraction submodule and a feature output submodule; a convolution operation is performed on the input feature to obtain a preprocessing feature; the preprocessing feature is input into the feature extraction submodule to obtain a first feature and a second feature; the first feature and the preprocessing feature are input into the pooling submodule to obtain a pooling feature; the pooling feature is input into the multi-scale extraction submodule to extract multi-scale information to obtain a multi-scale feature; the multi-scale feature, the preprocessing feature, the second feature and the input feature are input into the feature output submodule to obtain a spatial domain feature.

[0009] Furthermore, in the feature extraction submodule: at least two convolution operations of different scales are performed on the preprocessed features at the same time, and corresponding elements of the features output by the convolution operations are added to obtain a first feature, and the feature output by the convolution operation with the largest scale is the second feature.

[0010] Furthermore, in the pooling submodule: the first feature and the preprocessing feature are spliced; the spliced ​​features are subjected to average pooling and maximum pooling at the same time, and then the two pooled features are subjected to point-by-point convolution to obtain the pooled feature.

[0011] Furthermore, in the multi-scale extraction submodule: the pooled features are channel-separated, the separated features are convolved at different scales respectively, and then the convolved features are concatenated to obtain multi-scale features.

[0012] Furthermore, in the feature output submodule: a sigmoid activation operation is performed on the multi-scale features, the features after the sigmoid activation operation are multiplied by the corresponding elements of the preprocessing features and the second features respectively, and then the corresponding elements of the two features obtained by the multiplication are added to obtain the spatial domain weight; the spatial domain weight is multiplied by the corresponding element of the input feature to obtain the spatial domain feature.

[0013] Furthermore, in the frequency domain adaptive selection module: global feature pooling and fast Fourier transform are performed on the input features respectively; the features obtained by global feature pooling are subjected to multiple convolution operations and then subjected to sigmoid activation operations to obtain frequency domain weights; the frequency domain weights are multiplied by the corresponding elements of the static filter to obtain an adaptive filter; the features obtained by fast Fourier transform are subjected to the operation of the adaptive filter, and then the obtained features are subjected to inverse fast Fourier transform to obtain frequency domain features.

[0014] Furthermore, in the dual-domain feature interaction module: attention features are extracted for spatial domain features and frequency domain features respectively, and the second query matrix, the first key matrix and the first value matrix are obtained corresponding to the spatial domain features, and the first query matrix, the second key matrix and the second value matrix are obtained corresponding to the frequency domain features; the first query matrix, the first key matrix and the first value matrix are matrix multiplied, and then a convolution operation is performed after a Softmax operation to obtain a third feature; the second query matrix, the second key matrix and the second value matrix are matrix multiplied, and then a convolution operation is performed after a Softmax operation to obtain a fourth feature; after splicing the third feature and the fourth feature, the obtained spliced ​​features are respectively average pooled and maximum pooled, and the corresponding elements of the features obtained by the two pooling are added, and then convolution and sigmoid activation operations are performed in sequence to obtain the fusion weight; the fusion weight is multiplied by the corresponding element of the spliced ​​feature to obtain the fusion feature.

[0015] Furthermore, the target detection branch in step S2 includes multiple detection heads, and the number of detection heads is consistent with the number of initial features extracted by the feature extraction branch. Each detection head processes the fused features of corresponding sizes to obtain the category and position corresponding to the remote sensing image, and then combines the categories and all positions corresponding to the fused features of different sizes to determine the category and position corresponding to the remote sensing image; each detection head includes a category detection submodule and a position detection submodule; the fused features of corresponding sizes are input into the category detection submodule and the position detection submodule respectively, the category detection submodule determines the category corresponding to the current fused feature, and the position detection submodule determines the position corresponding to the current fused feature.

[0016] Compared with the prior art, the invention can achieve the following beneficial effects: (1) In the remote sensing image directional target detection method based on dual-domain feature fusion created by the present invention, the complementary features between different domains are fully utilized through adaptive selection of the spatial domain and the frequency domain and interactive fusion of features between the two domains, which effectively enhances the fusion of global context information and local information and makes up for the lack of target information; (2) In the remote sensing image directional target detection method based on dual-domain feature fusion created by the present invention, the proposed spatial domain adaptive selection module and frequency domain adaptive selection module are used to improve the target discrimination by enabling the network model to adaptively extract local and contextual information according to the characteristics of the target in the spatial domain and frequency domain; the proposed dual-domain feature interaction module is used to obtain complementary features by fusing the spatial domain and frequency domain features, thereby making up for the shortcomings of the single spatial domain features and enhancing the information expression capability of the target. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The drawings constituting part of the present invention are used to provide a further understanding of the present invention. The exemplary embodiments and descriptions of the present invention are used to explain the present invention and do not constitute an improper limitation on the present invention. In the drawings: Figure 1 A flowchart of a method for detecting directional targets in remote sensing images based on dual-domain feature fusion according to an embodiment of the present invention; Figure 2 A schematic diagram of a remote sensing target detection network according to an embodiment of the present invention; Figure 3 A schematic diagram of a spatial adaptive selection module according to an embodiment of the present invention; Figure 4 A schematic diagram of a frequency domain adaptive selection module according to an embodiment of the present invention; Figure 5 A schematic diagram of a dual-domain feature interaction module according to an embodiment of the present invention; Figure 6 A schematic diagram of a detection head according to an embodiment of the present invention. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solution and advantages of the invention more clear, the invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described here are only used to explain the invention and do not constitute a limitation of the invention.

[0019] 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. In the description of the present invention, it should be understood that the terms "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, the features defined as "first", "second", etc. may explicitly or implicitly include one or more of the features. In the description of the present invention, unless otherwise specified, the meaning of "multiple" is two or more.

[0020] The present invention will be described in detail below with reference to the accompanying drawings and in combination with embodiments.

[0021] like Figure 1 to Figure 2 As shown, the remote sensing image directional target detection method based on dual-domain feature fusion described in the embodiment of the present invention includes: S1: Obtain a remote sensing image dataset and preprocess the remote sensing image dataset to obtain a training set.

[0022] In one embodiment, the remote sensing image dataset uses the HRSC2016 dataset published at https: / / opendatalab.org.cn / OpenDataLab / HRSC2016 and the DIOR-R dataset published at https: / / opendatalab.org.cn / OpenDataLab / DIOR. The preprocessing comprises performing random flipping and random rotation operations on the remote sensing image dataset, with probabilities of 0.75 and 0.25 respectively.

[0023] S2: Construct a remote sensing target detection network. The remote sensing target detection network includes a feature extraction branch, a dual-domain feature fusion branch, and a target detection branch. The feature extraction branch extracts initial features of different sizes from the input remote sensing image; the dual-domain feature fusion branch extracts corresponding spatial and frequency domain features from initial features of different sizes, and fuses the spatial and frequency domain features of corresponding sizes to obtain corresponding fusion features; the target detection branch determines the corresponding category and position of the remote sensing image based on different fusion features.

[0024] In a specific embodiment, the feature extraction branch uses the ResNet50 network (published in the paper "Deep Residual Learning for Image Recognition" on the arXiv platform) to extract initial features, and the sizes of the extracted initial features are 1 / 4, 1 / 8, 1 / 16 and 1 / 32 of the input remote sensing image size. Four feature map scales of different sizes constitute the input of the dual-domain feature fusion branch.

[0025] In some embodiments, the dual-domain feature fusion branch includes a spatial domain adaptive selection module, a frequency domain adaptive selection module, a dual-domain feature interaction module, and a feature pyramid module. Among them, the spatial domain adaptive selection module extracts the corresponding spatial domain features from the initial features; the frequency domain adaptive selection module extracts the corresponding frequency domain features from the initial features; the dual-domain feature interaction module fuses the spatial domain features and frequency domain features of corresponding sizes to obtain the corresponding preliminary fused features; the feature pyramid module extracts multi-scale information from the preliminary fused features to obtain fused features. In a certain embodiment, the feature pyramid module adopts the FPN feature pyramid structure disclosed in the paper "Feature Pyramid Networks for Object Detection" from the arXiv platform.

[0026] In some embodiments, the structure of the spatial adaptive selection module is as follows: Figure 3 As shown, it includes a feature extraction submodule, a pooling submodule, a multi-scale extraction submodule and a feature output submodule. First, a convolution operation is performed on the input feature to obtain a preprocessing feature. The preprocessing feature is input into the feature extraction submodule to obtain a first feature and a second feature. The first feature and the preprocessing feature are input into the pooling submodule to obtain a pooling feature. The pooling feature is input into the multi-scale extraction submodule to extract multi-scale information to obtain a multi-scale feature. The multi-scale feature, the preprocessing feature, the second feature and the input feature are input into the feature output submodule together to obtain a spatial domain feature.

[0027] Specifically, in the feature extraction submodule, the preprocessing feature is simultaneously subjected to convolution operations of different scales for no less than 2 times, and the corresponding elements of the features output by the convolution operation are added to obtain the first feature, and the feature output by the convolution operation with the largest scale is the second feature. In the pooling submodule, the first feature and the preprocessing feature are spliced, the spliced ​​features are subjected to average pooling and maximum pooling at the same time, the two pooled features are spliced ​​through channels, and then the spliced ​​features are subjected to point-by-point convolution operations to obtain pooled features. In the multi-scale extraction submodule, the pooled features are channel-separated, the separated features are subjected to convolution operations of different scales respectively, and then the convolution features are spliced ​​to obtain multi-scale features. In the feature output submodule, the multi-scale features are subjected to sigmoid activation operations, the features after the sigmoid activation operations are respectively multiplied by the corresponding elements of the preprocessing feature and the second feature, and then the corresponding elements of the two features obtained by multiplication are added to obtain the spatial weight; the spatial weight is multiplied by the corresponding elements of the input feature to obtain the spatial feature.

[0028] In a specific embodiment, if Figure 3As shown in the figure, all convolution operations except the point-by-point convolution operation are depth-separable convolution operations, and the input features are subjected to depth-separable convolution operations with a convolution kernel size of 3×3 to obtain preprocessed features. The feature extraction submodule includes 4 depth-separable convolution operations of different scales, and the convolution kernel sizes are 5×5, 7×7, 9×9 and 11×11 respectively. At this time, it can be understood that the second feature is the feature obtained by the depth-separable convolution operation with a convolution kernel size of 11×11. The multi-scale extraction submodule includes 3 depth-separable convolution operations of different scales, and the convolution kernel sizes are 3×3, 1×11 and 11×1 respectively; channel separation is channel averaging, that is, the feature with a channel number of C is divided into 4 features with a channel number of C / 4, of which 3 features are subjected to 3 depth-separable convolution operations of different scales respectively, and the output features of the 3 depth-separable convolution operations of different scales are channel-joined with the remaining feature to obtain multi-scale features.

[0029] In some embodiments, the frequency domain adaptive selection module is as follows: Figure 4 As shown in the figure, global feature pooling and two-dimensional (2D) fast Fourier transform are performed on the input features respectively; the features obtained by global feature pooling are subjected to multiple convolution operations and then subjected to sigmoid activation operations to obtain frequency domain weights; the frequency domain weights are multiplied by the corresponding elements of the static filter to obtain an adaptive filter; the features obtained by fast Fourier transform are subjected to the operation of the adaptive filter, and then the obtained features are subjected to inverse fast Fourier transform to obtain frequency domain features.

[0030] It can be understood that the following two-dimensional fast Fourier transform is performed on the input features: ; Where x[h,w] represents the input feature, h and w are the ordinate and abscissa of each element in the input feature, H and W are the height and width of the input feature, respectively. X[h',w'] represents the feature obtained by fast Fourier transform, h' and w' are the ordinate and abscissa of each element in the feature obtained by fast Fourier transform, respectively.

[0031] After the features obtained by the two-dimensional fast Fourier transform are operated by the adaptive filter, the features obtained are subjected to the inverse fast Fourier transform as follows: ; Among them, X'[h',w'] represents the features obtained by the operation of the adaptive filter, and x'[h,w] represents the frequency domain features.

[0032] In a specific embodiment, global feature pooling is completed by global average pooling, and the features output by global feature pooling are continuously subjected to two two-dimensional (2D) convolution operations, wherein the features after the first two-dimensional convolution operation are subjected to a ReLU activation operation. It is understandable that the features after the second two-dimensional convolution operation are subjected to a sigmoid activation operation to obtain frequency domain weights. The static filter is a coefficient matrix composed of random numbers, and it is understandable that the adaptive filter is a coefficient matrix with frequency domain weights. The essence of the filtering operation of the adaptive filter is to perform a convolution operation on the features obtained by the fast Fourier transform using a coefficient matrix with frequency domain weights.

[0033] In some embodiments, the dual-domain feature interaction module is as follows Figure 5 As shown, attention features are extracted from spatial domain features and frequency domain features respectively, and the second query matrix, the first key matrix and the first value matrix are obtained corresponding to the spatial domain features, and the first query matrix, the second key matrix and the second value matrix are obtained corresponding to the frequency domain features; the first query matrix, the first key matrix and the first value matrix are matrix multiplied, and then a convolution operation is performed after a Softmax operation to obtain a third feature; the second query matrix, the second key matrix and the second value matrix are matrix multiplied and then a convolution operation is performed to obtain a fourth feature; after splicing the third feature and the fourth feature, the obtained spliced ​​features are respectively average pooled and maximum pooled, and the corresponding elements of the features obtained by the two poolings are added, and then convolution and sigmoid activation operations are performed in sequence to obtain the fusion weight; the fusion weight is multiplied by the corresponding element of the splicing feature to obtain the fusion feature.

[0034] It can be understood that attention features are extracted for spatial domain features and frequency domain features respectively. The spatial domain features correspond to the second query matrix Q2, the first key matrix K1 and the first value matrix V1, and the frequency domain features correspond to the first query matrix Q1, the second key matrix K2 and the second value matrix V2.

[0035] In a specific embodiment, the spatial domain features and the frequency domain features are first layer normalized and then the attention features are extracted. The first query matrix Q1, the first key matrix K1 and the first value matrix V1 are matrix multiplied, and then a two-dimensional convolution operation is performed after a Softmax operation to obtain the third feature, as shown in the following formula: ; Among them, F represents the third feature, C, M and N represent the number of channels, height and width of the spatial feature respectively, and Conv2d represents a two-dimensional convolution operation.

[0036] The second query matrix Q2, the second key matrix K1 and the second value matrix V2 are matrix multiplied, and then a two-dimensional convolution operation is performed after the Softmax operation to obtain the fourth feature, as shown in the above formula, which will not be repeated here.

[0037] After concatenating the third feature and the fourth feature, the concatenated features are respectively subjected to average pooling and maximum pooling. The corresponding elements of the features obtained by the two pooling are added together, and then one-dimensional convolution and sigmoid activation operations are performed in sequence to obtain the fusion weight; the fusion weight is multiplied by the corresponding element of the concatenated feature to obtain the fusion feature.

[0038] In some embodiments, the target detection branch in step S2 includes multiple detection heads, and the number of detection heads is consistent with the number of initial features extracted by the feature extraction branch. Each detection head processes the fusion features of corresponding sizes to obtain the category and position corresponding to the remote sensing image, and then combines the categories and all positions corresponding to the fusion features of different sizes to determine the category and position corresponding to the remote sensing image. The position specifically includes the horizontal and vertical coordinates of the target center point, the length and width of the target, and the rotation angle of the target. Each detection head includes a category detection submodule and a position detection submodule; the fusion features of corresponding sizes are input into the category detection submodule and the position detection submodule respectively. The category detection submodule determines the category corresponding to the current fusion feature, and the position detection submodule determines the position corresponding to the current fusion feature.

[0039] In a certain embodiment, the number of detection heads is 4, corresponding to features of 1 / 4, 1 / 8, 1 / 16 and 1 / 32 sizes respectively. Each detection head includes a category detection submodule and a position detection submodule: the category detection submodule includes 5 cascaded 3×3 two-dimensional convolution operations, the feature size and input features after the first four convolution operations remain unchanged, and the feature size after the last convolution operation is H×W×K (K is the number of categories); the position detection submodule also includes 5 3×3 two-dimensional convolution operations, the feature size and input features after the first four convolution operations remain unchanged, but the feature size after the last convolution operation is H×W×5. If the detection probabilities obtained by the four detection heads are within the preset confidence range, the corresponding categories and positions are retained, and then the redundant target frames are removed through the NMS (non-maximum suppression) processing operation. At this time, the category in the retained target frame is the category corresponding to the remote sensing image, and the position of the content in the retained target frame is the position corresponding to the remote sensing image. The setting of the confidence level is adjusted according to the actual situation, and is set to 0.25 in a certain embodiment.

[0040] S3: Use the training set obtained in step S1 to train the remote sensing target detection network in step S2 to obtain a remote sensing target detection model.

[0041] In one embodiment, the SGD optimizer is used, and the weight decay is 0.0001; the initial learning rate is 0.005, and it decreases by 10% in the 8th and 11th training cycles; for the HRSC2016 dataset, the Epoch is set to 36 rounds and the batch_size is set to 2; for the DIOR-R dataset, the Epoch is set to 12 rounds; the batch_size is set to 2; the position loss function is the Smooth L1 loss function, and the category loss function is the cross-entropy loss function.

[0042] S4: Input the remote sensing image to be detected into the remote sensing target detection model of step S3, and output the corresponding detection category and position.

[0043] In one embodiment, map50 (when the intersection-over-union ratio threshold is 0.5, when the overlap between the detection frame and the real frame reaches a certain degree, the correct proportion of the detection results, that is, the accuracy of the model detection results) is used as the evaluation index, and the target detection accuracy of the remote sensing images to be detected in the HRSC2016 data set is 90.7%, and the target detection accuracy of the remote sensing images to be detected in the DIOR-R data set is 65.7%. It can be seen that the target detection method provided by the present invention has a good accuracy rate.

[0044] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps described in the disclosure of the present invention can be performed in parallel, sequentially or in different orders, as long as the desired results of the technical solution disclosed in the present invention can be achieved, and this document does not limit this.

[0045] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement 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 directed targets in remote sensing images based on dual-domain feature fusion, characterized in that: include: S1: Acquire a remote sensing image dataset, and preprocess the remote sensing image dataset to obtain a training set; S2: Constructing a remote sensing target detection network; wherein the remote sensing target detection network includes: The feature extraction branch extracts initial features of different sizes from the input remote sensing image; The dual-domain feature fusion branch extracts the corresponding spatial domain features and frequency domain features from the initial features of different sizes; the spatial domain features and frequency domain features of the corresponding sizes are fused to obtain the corresponding fused features; The target detection branch determines the category and position corresponding to the remote sensing image according to different fusion features; S3: Using the training set obtained in step S1 to train the remote sensing target detection network in step S2, to obtain a remote sensing target detection model; S4: Input the remote sensing image to be detected into the remote sensing target detection model of step S3, and output the corresponding detection category and position.

2. The method for directional target detection in remote sensing images based on dual-domain feature fusion according to claim 1, characterized in that: The dual-domain feature fusion branch in step S2 includes: A spatial domain adaptive selection module extracts corresponding spatial domain features from the initial features; A frequency domain adaptive selection module extracts corresponding frequency domain features from the initial features; The dual-domain feature interaction module fuses the spatial domain features and frequency domain features of corresponding sizes to obtain the corresponding preliminary fusion features; The feature pyramid module extracts multi-scale information from the preliminary fusion features to obtain the fusion features.

3. The method for directional target detection in remote sensing images based on dual-domain feature fusion according to claim 2, characterized in that: The spatial domain adaptive selection module includes a feature extraction submodule, a pooling submodule, a multi-scale extraction submodule and a feature output submodule; Perform convolution operation on the input features to obtain preprocessed features; Inputting the preprocessing feature into the feature extraction submodule to obtain a first feature and a second feature; Inputting the first feature and the preprocessing feature into the pooling submodule to obtain a pooling feature; Inputting the pooled features into the multi-scale extraction submodule to extract multi-scale information to obtain multi-scale features; The multi-scale feature, the preprocessing feature, the second feature and the input feature are input into the feature output submodule to obtain the spatial domain feature.

4. The method for directional target detection in remote sensing images based on dual-domain feature fusion according to claim 3 is characterized in that: In the feature extraction submodule: The preprocessing features are simultaneously subjected to at least two convolution operations of different scales, and corresponding elements of the features output by the convolution operations are added to obtain the first feature, and the feature output by the convolution operation with the largest scale is the second feature.

5. The method for directional target detection in remote sensing images based on dual-domain feature fusion according to claim 3, characterized in that: In the pooling submodule: The first feature and the preprocessed feature are concatenated; the concatenated features are subjected to average pooling and maximum pooling at the same time, and then the two pooled features are subjected to point-by-point convolution to obtain the pooled feature.

6. The method for directional target detection in remote sensing images based on dual-domain feature fusion according to claim 3, characterized in that: In the multi-scale extraction submodule: The pooled features are channel-separated, the separated features are convolved at different scales, and the convolved features are concatenated to obtain the multi-scale features.

7. The method for directional target detection in remote sensing images based on dual-domain feature fusion according to claim 3, characterized in that: In the feature output submodule: A sigmoid activation operation is performed on the multi-scale feature, and the feature after the sigmoid activation operation is multiplied by the corresponding elements of the preprocessing feature and the second feature respectively, and then the corresponding elements of the two features obtained by the multiplication are added to obtain the spatial domain weight; the spatial domain weight is multiplied by the corresponding element of the input feature to obtain the spatial domain feature.

8. The method for directional target detection in remote sensing images based on dual-domain feature fusion according to claim 2, characterized in that: In the frequency domain adaptive selection module: Perform global feature pooling and fast Fourier transform on the input features respectively; The features obtained by the global feature pooling are subjected to multiple convolution operations and then subjected to a sigmoid activation operation to obtain frequency domain weights; the frequency domain weights are multiplied by the corresponding elements of the static filter to obtain an adaptive filter; After the features obtained by the fast Fourier transform are operated by the adaptive filter, the features obtained are subjected to inverse fast Fourier transform to obtain the frequency domain features.

9. The method for directional target detection in remote sensing images based on dual-domain feature fusion according to claim 2, characterized in that: In the dual-domain feature interaction module: Performing attention feature extraction on the spatial domain features and the frequency domain features respectively, the spatial domain features correspondingly obtain a second query matrix, a first key matrix and a first value matrix, and the frequency domain features correspondingly obtain a first query matrix, a second key matrix and a second value matrix; The first query matrix, the first key matrix and the first value matrix are matrix-multiplied, and then a convolution operation is performed after a Softmax operation to obtain a third feature; The second query matrix, the second key matrix and the second value matrix are matrix-multiplied, and then a convolution operation is performed after a Softmax operation to obtain a fourth feature; After splicing the third feature and the fourth feature, the obtained spliced ​​features are respectively subjected to average pooling and maximum pooling, and corresponding elements of the features obtained by the two poolings are added, and then convolution and sigmoid activation operations are performed in sequence to obtain fusion weights; the fusion weights are multiplied by the corresponding elements of the spliced ​​features to obtain the fusion features.

10. The method for directional target detection in remote sensing images based on dual-domain feature fusion according to claim 1, characterized in that: The target detection branch in step S2 includes a plurality of detection heads, and the number of the detection heads is consistent with the number of initial features extracted by the feature extraction branch. Each detection head processes the fused features of the corresponding size to obtain the category and position corresponding to the remote sensing image, and then combines the categories and all positions corresponding to the fused features of different sizes to determine the category and position corresponding to the remote sensing image. Each detection head includes a category detection submodule and a position detection submodule; The fused features of corresponding sizes are respectively input into the category detection submodule and the position detection submodule. The category detection submodule determines the category corresponding to the current fused feature, and the position detection submodule determines the position corresponding to the current fused feature.

Images