Angle attention-based anchor-frame-free remote sensing image rotating target detection method

Abstract

The invention provides an anchor-frame-free remote sensing image rotating target detection method based on angle attention, and belongs to the technical field of computer vision and remote sensing image processing. The method comprises the following steps: inputting a remote sensing image and extracting multi-scale features by using a backbone network; an optimal angle feature is extracted and selected through an angle feature extraction module and a feature pyramid network; a detection head is added to multi-scale output of the feature pyramid network for classification, frame regression and angle parameter regression; a frequency domain channel attention mechanism is introduced into an angle regression branch through an angle prediction head module in a detection head, so that the angle of a rotating target is predicted more accurately; and carrying out joint training by utilizing ProbIoU loss and distributed focus loss to realize end-to-end high-precision rotating target detection. According to the method, relatively high detection precision and reasoning speed are realized on a plurality of public remote sensing image data sets, and the method has efficient performance.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of computer vision and remote sensing image processing, and more particularly to a kind of anchor-free remote sensing image rotating target detection method based on angle attention. BACKGROUND

[0002] Aerial image target detection aims to assign a predefined class label to each instance in the image and determine its location. In recent years, with the continuous development of computer vision and remote sensing technology, the accuracy of remote sensing image target detection has been continuously improved. Unlike common target detection, the accuracy of angle regression has become a major problem faced by such detection models.

[0003] Currently, for natural scene images, target detection methods can be divided into two categories: anchor and anchor-free. The first category is "anchor-free" algorithms such as DRN and FCOSR, which are usually optimized for faster inference speed. The other category is "anchor" algorithm, S2A-Net and DA-Net belong to typical anchor method, which has achieved excellent performance on a series of benchmarks. However, current target detection algorithms still face some problems, especially aerial image rotating target detection in angle regression, which has a variety of problems that will lead to the inaccuracy of the detection results. Specifically, these problems mainly include the following three points: (1) Arbitrary direction. In remote sensing images, targets may be arranged in any direction, which puts higher requirements on the angle prediction accuracy of deep learning models. Traditional bounding box methods often have difficulty accurately capturing these rotating targets, resulting in insufficiently accurate angle prediction. Since targets such as airplanes, ships, and vehicles may be positioned in different directions, conventional horizontal or vertical bounding boxes cannot effectively cover their true shape and direction, thereby affecting the accuracy of subsequent classification and identification tasks.

[0004] (2) Large aspect ratio. In remote sensing images, the size and shape of targets are extremely diverse, and the difference between object classes and spatial resolution leads to significant differences in target size. This size variation is particularly challenging for traditional convolutional neural networks, as these networks typically use fixed receptive fields to process images. The limitations of fixed receptive fields mean that the network may not be able to effectively capture the details of extremely small or large objects, thereby affecting the accuracy and reliability of target detection. For example, for small targets, a fixed receptive field may cover too much background information, making feature extraction less focused; for large targets, important features may be scattered across multiple receptive fields, making it difficult to be captured as a whole.

[0005] (3) Dense arrangement. Remote sensing images captured by satellites usually have large sizes and contain a large number of objects. In this case, the dense arrangement and mutual occlusion between targets not only limit the features extracted from the convolutional layer, but also increase the difficulty of target recognition and positioning. When multiple targets are very close or partially overlapped, traditional detection algorithms can mistakenly identify them as a single entity or completely miss partially occluded targets. In addition, the dense arrangement of the environment can also cause the algorithm to misjudge the background features as targets, further reducing the accuracy of detection.

[0006] In addition, with the development of target detection technology, many anchor-based detection architectures have been applied to aerial images. Anchor-based rotated target detection can be divided into one-stage and two-stage methods. For two-stage networks, dozens of anchors with different scales, angles and aspect ratios are usually used to reduce the problems caused by irregularly oriented and large aspect ratio targets in aerial images. S2A-Net and DA-Net, as one-stage detectors, better extract effective information from densely arranged targets by aligning instance-level and image-level features. However, due to the need to set anchors with multiple aspect ratios, angles and scales in the network, these methods are computationally complex and have high memory overhead, making model training difficult. SUMMARY

[0007] To solve the above problems, the purpose of the present application is to provide an anchor-free rotated target detection method for remote sensing images based on angle attention, which solves the problems of insufficient angle regression accuracy and high computational complexity in existing rotated target detection technology by constructing a network structure combining multi-scale feature extraction and angle attention modules.

[0008] To achieve the above purpose, the technical scheme is as follows: In a first aspect, the present application provides an anchor-free rotated target detection method for remote sensing images based on angle attention, comprising: inputting a remote sensing image and extracting multi-scale features using a pre-trained backbone network; adding an angle feature extraction module between the backbone network and the feature pyramid network, using the angle feature extraction module to generate multi-scale features in parallel through depth separable strip convolution after dimension reduction of the multi-scale features, and combining a rotation matrix to perform angle selection and resampling on the multi-scale features to generate enhanced multi-scale features; inputting the enhanced multi-scale features into the feature pyramid network, outputting target multi-scale features in a five-layer pyramid structure through horizontal connection, layer-by-layer upsampling and convolution fusion operations; respectively connecting detection heads on the multi-scale outputs of the feature pyramid network for classification, bounding box regression and angle parameter regression; In the detection head, an angle prediction head module is introduced in the angle regression branch, a frequency domain channel attention mechanism is introduced, target multi-scale features are mapped to the frequency domain through Fourier transform, and the angle is independently regressed based on integral operation after channel attention is used to recover to the spatial domain through inverse Fourier transform; The regression information is optimized by ProbIoU loss and distributed focal loss to generate the detection result.

[0009] In an optional implementation, the input remote sensing image is extracted using a pre-trained backbone network to obtain multi-scale features, including: The remote sensing image is inputted into a pre-trained ResNet-50 backbone network; The ResNet-50 backbone network includes a convolutional layer, a pooling layer, four-stage residual modules and a bottleneck structure. The ResNet-50 backbone network includes a convolutional layer, a pooling layer, four-stage residual modules and a bottleneck structure. The ResNet-50 backbone network includes a convolutional layer, a pooling layer, four-stage residual modules and a bottleneck structure.

[0010] In an optional implementation, the multi-scale features are generated in parallel through deep separable bar convolution after dimension reduction by the angle feature extraction module, and the multi-scale features are angle-selected and resampled by combining a rotation matrix to generate enhanced multi-scale features, including: The input multi-scale features are extracted by the angle feature extraction module; The feature extraction module includes a 1x1 convolutional layer, a multi-branch convolutional structure and a channel attention mechanism. H and W represent the number of channels, spatial height and width, respectively, C3, C4 and C5 represent the first, second and third groups of feature maps output by the ResNet-50 backbone network, and the value range is 3, 4 and 5. The input multi-scale features are reduced in channel number through a 1x1 convolutional layer to obtain a compact representation . The context information of is captured through a multi-branch convolutional structure, and the relationship between different channels is modeled. The output of the 1x1 convolution is directly used as the attention weight to weight . The angle feature extraction process is represented as follows:

[0011]

[0012] Wherein and are the attention map and output of feature extraction, respectively, is an element-wise matrix multiplication operation, and DW_Conv represents a depth convolution. The first branch of the multi-branch convolution structure is represented by scale0 is a residual connection.

[0013] In an optional implementation, the enhanced multi-scale feature is input into the feature pyramid network, and target multi-scale features in a five-layer pyramid structure are output through horizontal connection, layer-by-layer upsampling and convolution fusion operation, including: The enhanced multi-scale feature is denoted as and is input into the feature pyramid network. 1x1 convolution dimension reduction is performed on, after the number of channels is unified, layer-by-layer upsampling is performed from top to bottom, and element-by-element addition is performed with the features of the next layer, and then 3x3 convolution is performed for smoothing processing, and finally pyramid features P3, P4 and P5 are obtained. Layer-by-layer downsampling is performed on the feature map of P5 through 3x3 convolution, and pyramid features P6 and P7 are obtained, which can be formally represented as:

[0014] wherein, represents the output feature map of the layer FPN, is the output feature map of the corresponding layer in the angle feature extraction module, Upsample(·) represents the upsampling operation, the value range of k is 3, 4 or 5. Based on the pyramid features P3, P4, P5, P6 and P7, target multi-scale features are generated.

[0015] In an optional implementation, detection heads are respectively connected on the multi-scale outputs of the feature pyramid network, for classification, bounding box regression and angle parameter regression, including: Pyramid features P3, P4, P5, P6 and P7 output by the feature pyramid network are respectively provided with detection head branches; the detection head branches include parallel classification branches, bounding box regression branches and direction branches; The classification branch is used to determine whether there is a target at each position and give the class probability thereof. The bounding box regression branch is used to predict the center position coordinates (x, y) and scale size (w, h) parameters of the target bounding box; wherein w is the width of the target bounding box and h is the height of the target bounding box. The direction branch is used to predict the rotation angle parameter θ of the target.

[0016] In an optional implementation, the frequency domain channel attention mechanism includes: The attention of the channel is adjusted by the following formula:

[0017] wherein, represents a sigmoid function, represents a full link layer, represents the obtained multi-spectrum vector.

[0018] In an optional embodiment, the regression information is optimized by using the ProbIoU loss and the distributed focal loss, comprising: The regression information including target position (x, y), size (w, h) and rotation angle θ is optimized by using the ProbIoU loss, and the rotation angle θ is predicted by using the distributed focal loss; The form of the ProbIoU loss is:

[0019] wherein, H i represents the probability distribution value of the predicted bounding box at the pixel point i; G i represents the probability distribution value of the real bounding box at the pixel point i; A smoothing term is used to prevent numerical instability; The form of the distributed focal loss is:

[0020] wherein, represents the probability of the angle decreasing in each interval, is set to π / 180.

[0021] In a second aspect, the embodiments of the present application also provide an anchor-free remote sensing image rotation target detection system based on angle attention, comprising: A first feature extraction module is configured to input a remote sensing image and extract multi-scale features by using a pre-trained backbone network; A second feature extraction module is configured to add an angle feature extraction module between the backbone network and the feature pyramid network, use the angle feature extraction module to generate multi-scale features by parallelly passing through a depth separable bar convolution after dimension reduction of the multi-scale features, and combine a rotation matrix to perform angle selection and resampling on the multi-scale features to generate enhanced multi-scale features; A third feature extraction module is configured to input the enhanced multi-scale features into the feature pyramid network, output target multi-scale features in a five-layer pyramid structure by horizontal connection, layer-by-layer upsampling and convolution fusion operations; A detection head setting module is configured to connect detection heads on the multi-scale outputs of the feature pyramid network respectively, for classification, bounding box regression and angle parameter regression; The detection module is configured to introduce a frequency domain channel attention mechanism in an angle regression branch in the detection head through the angle prediction head module, map target multi-scale features to a frequency domain through Fourier transform, restore to a spatial domain after channel attention, and independently regress a rotation angle based on integral operation. The optimization module is configured to jointly optimize regression information by using ProbIoU loss and distributed focal loss to generate a detection result.

[0022] In a third aspect, the embodiments of the present application further provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the steps of the angle attention-based anchor-free remote sensing image rotating target detection method according to any one of the above.

[0023] In a fourth aspect, the embodiments of the present application further provide a storage medium having a computer program stored thereon, and the computer program is executable by a processor to implement the steps of the angle attention-based anchor-free remote sensing image rotating target detection method according to any one of the above.

[0024] From the above technical solutions, the present application has the following advantages: In the angle attention-based anchor-free remote sensing image rotating target detection method provided by the present application, the design of combining the angle attention mechanism with the multi-scale feature extraction is combined, the angle feature extraction module and the angle regression head module are independently constructed, the interference of the angle and the position parameters is effectively decoupled, and the rotation angle regression accuracy is improved. When dealing with the rotating targets with multiple directions, high density, and large width-height ratio changes in the remote sensing image, the present application has stronger generalization ability and robustness.

[0025] The present application introduces the frequency domain channel attention mechanism in the angle regression process, and combines the integral regression method, further enhances the distinguishability and expressiveness of the angle features. Experiments show that, compared with the traditional angle regression method, the present application can significantly improve the detection accuracy. The ProbIoU loss and the distributed focal loss are used to jointly optimize the rotation box position and the angle parameter, improve the stability and convergence speed of the model in the rotating target detection task, and have strong practicality and popularization value.

[0026] The present application uses the pre-trained ResNet-50 backbone network to extract the multi-scale features of the remote sensing image, combines the transverse connection, up-sampling, and convolution fusion operation of the feature pyramid network, outputs the target multi-scale features in a five-layer pyramid structure, can fully capture the feature information of different size targets, effectively improves the detection ability of the targets with large scale changes in the remote sensing image, and improves the accuracy and comprehensiveness of the detection.

[0027] The application increases an angle feature extraction module between a backbone network and a feature pyramid network, generates enhanced multi-scale features through dimension reduction, depth separable bar convolution, angle selection and resampling; meanwhile, introduces a frequency domain channel attention mechanism in the detection head angle regression branch, and independently regresses the rotation angle based on integral operation. This series of operations effectively enhances the extraction and utilization of target rotation angle features, improves the detection accuracy of rotating targets, and is especially suitable for scenes with variable target directions in remote sensing images.

[0028] The application sets a detection head on the multi-scale output of the feature pyramid network, which contains classification, bounding box regression and angle parameter regression branches, and realizes comprehensive prediction of target categories, positions, scales and rotation angles. This design enables the model to more accurately locate and identify targets in remote sensing images, output more rich target information, and meet the demand for fine detection of targets in complex remote sensing scenes.

[0029] The application uses ProbIoU loss and distributed focal loss to jointly optimize the regression information, ProbIoU loss optimizes target position, size and rotation angle, and distributed focal loss specially predicts rotation angle. The combination of the two effectively improves the learning efficiency and accuracy of target regression information in the model training process, makes the model converge faster and the detection result more accurate, and enhances the generalization ability and robustness of the model. BRIEF DESCRIPTION OF DRAWINGS

[0030] In order to more clearly illustrate the technical solutions of the present application, the drawings needed in the description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0031] Figure 1 The flowchart of the angle attention based anchor-free remote sensing image rotating target detection method provided by the present application.

[0032] Figure 2 The overall network structure diagram of the angle attention based anchor-free remote sensing image rotating target detection method provided by the present application.

[0033] Figure 3 The structure diagram of the angle feature extraction module provided by the present application.

[0034] Figure 4 The structure diagram of the angle attention based anchor-free remote sensing image rotating target detection system provided by the present application.

[0035] Figure 5 The structure diagram of the electronic device provided by the present application. DETAILED DESCRIPTION

[0036] Various embodiments of the present disclosure will be described in detail below with reference to the detailed description of the angle attention-based anchor-free remote sensing image rotating target detection method. The present disclosure can have various embodiments, and adjustments and changes can be made therein. However, it should be understood that there is no intention to limit various embodiments of the present disclosure to the specific embodiments disclosed herein, but the present disclosure should be understood to cover all adjustments, equivalents and / or alternatives falling within the spirit and scope of various embodiments of the present disclosure.

[0037] The core of the present application is to provide an angle attention-based anchor-free remote sensing image rotating target detection method. In related art, rotating target detection has made significant progress in the field of remote sensing image processing. Unlike conventional target detection, the rotating target detection bounding box needs accurate angle parameters to locate the target. However, due to the arbitrary direction of the target in the aerial image, the large aspect ratio and the dense arrangement, the accuracy of angle regression is still a challenge to be solved.

[0038] The angle attention-based anchor-free remote sensing image rotating target detection method provided by the present application has the following basic idea: common targets (such as airplanes, ships, vehicles, etc.) in remote sensing images often have problems such as arbitrary direction, large size difference, and dense distribution. Traditional target detection methods are mostly based on anchor box mechanism, which requires predefining a large number of anchor box templates with different scales, aspect ratios and rotation angles. Not only is the design complex, but also the calculation and memory overhead is large, making it difficult to adapt to the diversity and complexity of actual scenes. Especially in rotating target detection, the angle regression and position parameters are coupled, and the traditional method often has the problems of unstable regression and decreased detection accuracy. In order to solve the above problems, the present application proposes an angle attention-based anchor-free remote sensing image rotating target detection method, which aims to realize independent modeling and fine regression of the rotating angle of the target without predefining anchor boxes, by using a specially designed angle feature extraction module and an angle prediction head module, so as to realize higher precision and higher efficiency of rotating target detection.

[0039] Specifically, the network first extracts multi-scale features through a ResNet-50 backbone network, and further integrates semantic information of different levels in combination with a feature pyramid network (FPN). Between the scale output by the backbone network and the input of the FPN, an angle feature extraction module is added to explicitly model the direction information through multi-scale bar convolution and rotation sampling mechanism. Subsequently, an angle prediction head module is introduced to realize independent regression of the angle with the aid of frequency domain channel attention, thereby solving the instability problem caused by the coupling of the angle and other position parameters. Finally, a joint loss function is used to train the position, scale and angle of the detection frame in an end-to-end manner, thereby improving the detection accuracy and convergence speed. The method of the present application does not rely on pre-defined anchor frames, significantly reduces the complexity of the network structure and the training cost, and at the same time, through the modularized angle perception design, effectively enhances the modeling capability of the rotation angle, improves the detection accuracy and speed, and is suitable for the rotation target detection task in various remote sensing scenes, including city aerial photography, port monitoring, etc.

[0040] Hereinafter, the term "include" or "may include" used in various embodiments of the present disclosure indicates the presence of the disclosed functions, operations, or elements, and does not limit the addition of one or more functions, operations, or elements. In addition, as used in various embodiments of the present disclosure, the terms "include", "have", and their synonyms merely mean to indicate the presence of a specific feature, number, step, operation, element, component, or combination of the foregoing, and should not be understood as first excluding the presence or addition of one or more other features, numbers, steps, operations, elements, components, or combinations of the foregoing.

[0041] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0042] Please refer to Figure 1 The method flowchart of a kind of anchor-free remote sensing image rotating target detection method based on angle attention in a specific embodiment is shown, and the method comprises the following steps: S1: input remote sensing image and extract multi-scale features using a pre-trained backbone network.

[0043] In the specific embodiment, the remote sensing image Input a pre-trained ResNet-50 backbone network; the ResNet-50 backbone network comprises a convolutional layer, a pooling layer, four-stage residual modules and a bottleneck structure; three layers of feature maps C3, C4 and C5 of different scales are extracted from the ResNet-50 backbone network in a remote sensing image to generate multi-scale features.

[0044] S2: An angle feature extraction module is added between the backbone network and the feature pyramid network, the multi-scale features are reduced in dimension by the angle feature extraction module, then the multi-scale features are generated in parallel through deep separable bar convolution, and the multi-scale features are angle-selected and resampled by a rotation matrix to generate enhanced multi-scale features.

[0045] In the specific embodiment, an angle feature extraction module (AFE) is added between the backbone network and the feature pyramid network (FPN), the AFE module generates multi-scale features through 1x1 convolution after dimension reduction, then the multi-scale features are generated in parallel through 3x1, 5x1, 7x1 and corresponding 1x3, 1x5, 1x7 deep separable bar convolution, and the multi-scale features are angle-selected and resampled by a rotation matrix to generate enhanced multi-scale features.

[0046] S3: The enhanced multi-scale features are input into the feature pyramid network, and target multi-scale features of a five-layer pyramid structure are output through horizontal connection, layer-by-layer upsampling and convolution fusion operations.

[0047] S4: A detection head is connected to the multi-scale output of the feature pyramid network respectively for classification, bounding box regression and angle parameter regression.

[0048] In the specific embodiment, a detection head is connected to the multi-scale output of the FPN respectively for classification, bounding box regression and angle parameter regression.

[0049] S5: In the detection head, a angle prediction head module is used to introduce a frequency domain channel attention mechanism in the angle regression branch, the target multi-scale features are mapped to the frequency domain through Fourier transform, are recovered to the spatial domain after channel attention using inverse Fourier transform, and the rotation angle is independently regressed based on integral operation.

[0050] In the specific embodiment, in the detection head, a angle prediction head (APH) module is used to introduce a frequency domain channel attention mechanism in the angle regression branch, the features are mapped to the frequency domain through Fourier transform, are recovered to the spatial domain after channel attention using inverse Fourier transform, and the rotation angle θ is independently regressed based on integral operation, which facilitates more accurate prediction of the angle of the rotating target.

[0051] S6: The regression information is optimized by using ProbIoU loss and distributed focal loss in combination to generate a detection result.

[0052] In this step, the ProbIoU loss and the distributed focus loss (DFL) are combined to optimize the regression information and the rotation angle θ, achieving end-to-end high-precision rotation target detection.

[0053] Based on the above method, as shown in Figure 2 The overall network structure of the method is an anchor-free network based on angle attention, including a backbone network, a feature pyramid network (FPN), and a detection head. This structure is entirely composed of convolutional layers, demonstrating its simplicity and efficiency. In this patent method, the FCOSR typical anchor-free one-stage target detector is improved. The FCOSR method introduces an ellipse center sampling method, reduces low-quality detection results, and significantly improves the overall performance of the model through this technology. In addition, in order to further optimize the network performance, the method also uses fuzzy sample label assignment and multi-level sampling strategy. Five parameters are used for regression of the predicted bounding box. The parameters represent the coordinates of the upper left corner of the bounding box, while and represent its width and height, respectively. In addition, represents the rotation angle of the bounding box. These five parameters have different ranges, and can be negative, but and must be positive, must be limited to This method takes FCOSR as the baseline and further introduces two core modules, the angle feature extraction module and the angle prediction head module, to improve the accuracy of rotation target detection. These new modules aim to more effectively handle the complex target directions and shapes commonly found in remote sensing images, thereby accurately regressing the angle of each target.

[0054] Based on the above network architecture, the specific process of the anchor-free box remote sensing image rotation target detection method based on angle attention disclosed in this embodiment is as follows: for a given remote sensing image input , First, it is input into the backbone network ResNet-50, and three layers of feature maps C3, C4 and C5 of different scales are extracted, corresponding to high resolution, medium resolution and low resolution spatial perception semantic information, forming a multi-scale basic feature representation. Next, the extracted multi-scale features are sent to the angle feature extraction module for angle feature extraction, which includes multi-scale depth separable bar convolution (such as 3x1, 5x1, 7x1, etc.) and rotation sampling mechanism, which transforms the receptive field direction through a rotation matrix, thereby enhancing the network's ability to model targets in any direction and extracting angle-aware features. The extracted angle features are input into the feature pyramid network (FPN), which outputs five layers of pyramid structure P3 to P7 through horizontal connection, layer-by-layer upsampling and convolution fusion operations, further enhancing the multi-scale fusion capability of the features. These feature maps are output as the backbone output, and the output of each layer of the FPN is input into the detection head. Then, three detection branches are set in parallel in each detection head for classification, regression and angle prediction. The classification branch outputs the class probability , where N is the number of classes; the regression branch predicts the center point position and scale parameter ; and the angle regression branch independently predicts the rotation angle parameter θ . Finally, the network outputs the rotated target detection result , where represents the class label, realizing high-precision positioning and recognition of targets in any direction in the remote sensing image.

[0055] In this embodiment, efficient multi-scale feature extraction is achieved by pre-training the backbone network combined with the feature pyramid network, angle feature extraction is strengthened by using the angle feature extraction module and the frequency domain channel attention mechanism, multi-information accurate prediction of the target is realized by optimizing the detection head, and the regression information is optimized by combining the ProbIoU loss and the distributed focal loss, thereby significantly improving the detection accuracy and comprehensiveness of the target with large scale changes and variable directions in the remote sensing image, enhancing the model generalization ability and robustness, and meeting the fine detection requirements of the target in complex remote sensing scenarios.

[0056] In an embodiment of the present application, based on step S1, a possible embodiment will be given below to non-restrictively describe the specific implementation scheme.

[0057] In this embodiment, the ResNet-50 backbone network includes convolutional layers, pooling layers, four residual stages, and a feature output part, and the specific implementation is as follows: inputting a remote sensing image , first performing preliminary feature extraction through 7x7 convolution, and the expression is as follows:

[0058] Where BN is the BatchNorm operation, and ReLU is the activation function.

[0059] Then down-sampling through the max-pooling layer:

[0060] After that, the image features are extracted layer by layer through four-stage residual modules, respectively , , and . Each stage contains several bottleneck structures. Each bottleneck structure includes a 1x1 convolution to reduce the dimension, a 3x3 convolution to process the main features, a 1x1 convolution to restore the dimension, a BatchNorm and a ReLU activation operation. Among them, the first stage and the fourth stage contain 3 bottleneck structures, the second stage contains 4 bottleneck structures, and the third stage contains 6 bottleneck structures.

[0061] In the four stages, the feature map size gradually decreases, and the channel number gradually increases, respectively outputting three groups of feature maps C3, C4 and C5:

[0062]

[0063]

[0064]

[0065] Therefore, the ResNet-50 backbone network extracts three layers of feature maps C3, C4 and C5 of different scales in the remote sensing image to generate multi-scale features.

[0066] In an embodiment of the present application, based on step S2, a possible embodiment will be given below to illustrate the specific implementation of the embodiment.

[0067] In this embodiment, the angle feature extraction module is shown in the schematic diagram as Figure 3 The angle feature extraction module includes angle feature extraction and angle encoding two parts, and the specific implementation is as follows: Given the input feature map , where , H, W represent the channel number, spatial height and width, respectively. represents the first group of feature maps output by the ResNet-50 backbone network, and its value range is 3, 4, 5. After a 1x1 convolution layer reduces the channel number, a compact representation Then, a multi-branch convolution structure (1x3 and 3x1 branches, 1x5 and 5x1 branches, 1x7 and 7x1 branches) is used to capture extensive context information and model the relationship between different channels. The output of 1x1 convolution is directly weighted as attention weight to the input. Specifically, the angle feature extraction process can be written as:

[0068]

[0069] where and are the attention map and output of feature extraction, respectively, and takes the value range of 3, 4, 5. is the element-wise matrix multiplication operation. DW_Conv represents the depthwise convolution, the th branch of the multi-branch convolution structure, and scale0 is the residual connection.

[0070] Next, the output angle feature is connected to the rotation convolution layer (RCL). Inspired by DCN, the grid R defines the receptive field size and dilation rate. For a 7x7 convolution with a dilation factor of 2, it can be represented as:

[0071] For the angle learned at the th position, is the rotation matrix, and the predefined offset , is defined as:

[0072] For each position on the output feature map , there is:

[0073] where enumerates the positions in R, , is the output of feature extraction, is the convolution kernel weight, ensuring accurate encoding and extraction of angle information.

[0074] Rotational convolutional layers (RCLs) introduce an angle-learnable rotational receptive field, enabling the convolutional kernel to adapt to different orientations of targets in an image. Compared to traditional convolutions that only perceive local regions in a fixed orientation, RCLs can dynamically adjust the rotation angle of the receptive field, thus more effectively capturing the structural features of rotating targets. This mechanism is particularly suitable for scenarios with varying target orientations, such as remote sensing images and text images. By combining with dilated convolutions, RCLs can also maintain high resolution while expanding the receptive field, effectively improving the detection accuracy and robustness of rotating targets.

[0075] In one embodiment of the present invention, based on step S3, the following will provide a possible embodiment and describe its specific implementation in a non-limiting manner.

[0076] In this embodiment, the Feature Pyramid Network (FPN) is used to generate multi-level semantic feature maps with different resolutions but a uniform number of channels. The FPN structure establishes semantic enhancement paths layer by layer from top to bottom, and achieves the fusion of high- and low-level features through lateral connections.

[0077] Specifically, the features output by the angle feature extraction module After performing 1×1 convolutional dimensionality reduction and unifying the number of channels, upsampling is performed layer by layer from top to bottom, and the features are added element-wise to the features of the next layer. Then, a 3×3 convolution is used for smoothing, ultimately yielding pyramid features P3, P4, and P5. To better extract high-dimensional features for detecting large-scale targets, the feature map of P5 is downsampled layer by layer using a 3×3 convolution to obtain features P6 and P7. This process can be formally represented as:

[0078] in, Indicates the first Layer FPN outputs feature maps. This represents the output feature map of the corresponding layer in the angle feature extraction module. Upsample(⋅) indicates the upsampling operation. The value range is 3, 4, 5.

[0079] In one embodiment of the present invention, based on step S4, the following will provide a possible embodiment and describe its specific implementation in a non-limiting manner.

[0080] In this embodiment, reference Figure 2 As shown, the detection head module contains three parallel detection branches: angle prediction, regression, and classification branches, which are used to predict the rotation angle, target position and size, and class probability, respectively. The three branches share the multi-scale features P3–P7 extracted by the backbone network and FPN, and perform predictions independently on the feature maps of each scale, forming a multi-scale detection structure.

[0081] In an embodiment of the present application, based on step S5, a possible embodiment will be given below to specifically and non-limitingly illustrate the specific implementation thereof.

[0082] The angle prediction branch inputs the enhanced feature map of the angle feature extraction module (AFE), performs channel compression through 1x1 convolution, and then inputs the angle prediction head module (APH) to model the frequency domain channel attention and angle integral regression, and outputs the rotation angle parameter θ∈[0°, 90°]. The bounding box regression branch is used to predict the center point coordinates (x, y) and the width and height parameters (w, h) of the rotating box. The input is the FPN feature map, which is output after a plurality of 3x3 convolution layers. The output is 4 channels of coordinate and scale regression values. The target classification branch is used to predict the probability distribution of each position belonging to a certain category. The input feature is extracted through the convolution layer to obtain the category-related feature, and finally the classification probability P∈[0, 1] of each category is output. The output results of the three detection branches are finally spliced into the rotating box detection results: .

[0083] Wherein, c is the category label predicted by the classification branch. The three-branch design of the detection head has the characteristics of mutual non-interference and independent optimization, effectively solving the training instability problem caused by the coupling of angle and position parameters, and improving the precision and robustness of the rotating target detection.

[0084] It should be particularly noted that the frequency domain channel attention (FCA) uses the frequency domain to enhance the attention effect, which is particularly effective for aerial images, because the frequency attention mechanism can represent the channel information in a rich manner. The obtained multi-spectrum vector is represented as follows: And the attention of the channel is adjusted through the following formula:

[0085] Wherein represents the sigmoid function, represents the full-link layer, represents the obtained multi-spectrum vector. This design not only strengthens the detection ability of the model for the rotating target, but also makes the angle prediction more accurate.

[0086] In an embodiment of the present application, based on step S6, a possible embodiment will be given below to specifically and non-limitingly illustrate the specific implementation thereof.

[0087] The ProbIoU loss is used to optimize the overall regression information The core idea of the ProbIoU loss is to express the parameters of the predicted box and the real box as a probability distribution, and then calculate the IoU according to the overlap degree between the distributions to improve the accuracy of the bounding box positioning and the rotation regression. Its expression form is:

[0088]

[0089] wherein, Hi represents the probability distribution value of the predicted bounding box at the pixel point Gi; Gi represents the probability distribution value of the real bounding box at the pixel point Gi; To prevent numerical instability of the smoothing term.

[0090] In order to calculate the ProbIoU loss, it is necessary to convert the rotated bounding box into a Gaussian bounding box. Due to the problem of being unable to determine the direction due to the existence of the Gaussian bounding box, the Gaussian bounding box can be a square rotated bounding box. In order to solve this problem, a distribution focal loss (DFL) is adopted to predict the rotation angle of the bounding box. Specifically, the integral form of the predefined wherein, is an even interval of angular discretization. The formula is as follows:

[0091] wherein, represents the probability of the angle falling within each interval, is set to π / 180.

[0092] Further, in order to verify the effectiveness of the angle attention-based anchor-free remote sensing image rotating target detection method disclosed in the present application, the detection effect of the model is evaluated on special remote sensing benchmark data sets such as the DOTA data set and the HRSC2016 data set. On the above data sets, the mean average precision (Mean Average Precision, mAP) and FPS are used as two indicators for quantitative evaluation. The higher the value of the indicator, the better the performance of the model. Specifically, the mean average precision (mAP) is the average value of the detection performance of all classes by calculating the intersection over union (Intersection over Union, IoU) between the predicted box and the real box, and is usually used to evaluate the performance of the entire model in multi-class target detection. FPS represents the number of frames that the model can process per second in the inference stage, which is an important indicator for measuring the real-time performance of the target detection system.

[0093] In the present embodiment, in order to realize training and testing, the data organization mode of training set, validation set and test set is adopted. All images are uniformly adjusted to 800x800 pixels for easy processing while maintaining the original aspect ratio of the images. Such processing not only facilitates network learning, but also helps to maintain the visual information of the images, thereby ensuring the detection accuracy.

[0094] In this embodiment, when conducting experiments, a ResNet-50 pre-trained on ImageNet is selected as the backbone network of the model. For the DOTA dataset, the model is trained for a total of 36 training cycles; while for the HRSC2016 dataset, the training process is completed through 40,000 iterations. When training the DOTA dataset, the SGD optimizer with an initial learning rate of 0.01 is used, and the learning rate is reduced by 10 times at the 24th and 33rd training cycles. For the HRSC2016 dataset, the initial learning rate is set to 0.001, and step decay is set at 30,000 and 36,000 iterations. The model training uses a weight decay of 0.0001 and a momentum of 0.9. All training and inference operations are performed on a computing device equipped with 4 NVIDIA 4090 GPUs with 24 GB of memory, and the overall batch size is set to 4.

[0095] The angle attention-based anchor-free remote sensing image rotation target detection method disclosed in the present application is compared with eight state-of-the-art representative methods, including PIoU, S2A-Net, ProbIoU, DRN, CenterMap, BBAVectors, PolarDet, and FCOSR-L. The quantitative comparison results are shown in Table 1.

[0096] Table 1: Comparison of objective evaluation results of different detection methods on the HRSC2016 dataset

[0097] In Table 1, mAP 12 and 07 represent the VOC2012 and VOC2017 indicators, respectively. FPS represents the number of frames per second.

[0098] In the HRSC2016 dataset, targets usually have arbitrary directions and large aspect ratios. Anchor-based target detection methods, such as S2A-Net, achieve better detection results by setting more anchors, but this approach often sacrifices speed. The anchor-free method proposed in the present application achieves an mAP of 95.96% on this dataset while maintaining a processing speed of 19.7 FPS. In addition, the present application is tested on the VOC2007 and VOC2012 evaluation standard datasets, verifying the universality of the method. In the VOC2012 evaluation, the FAA-Net method proposed in the present application is 1.12% higher in mAP than the FCOSR-L method, proving the effectiveness and performance of the method.

[0099] As Figure 4As shown, the following is an embodiment of an angle attention-based anchor-free remote sensing image rotating target detection system provided by the embodiment of the present disclosure. The system and the angle attention-based anchor-free remote sensing image rotating target detection method of each of the above embodiments belong to the same inventive concept. Details not described in the embodiment of the angle attention-based anchor-free remote sensing image rotating target detection system can be referred to the embodiment of the angle attention-based anchor-free remote sensing image rotating target detection method described above.

[0100] An angle attention-based anchor-free remote sensing image rotating target detection system comprises: A first feature extraction module is configured to input a remote sensing image and extract multi-scale features using a pre-trained backbone network.

[0101] A second feature extraction module is configured to add an angle feature extraction module between the backbone network and the feature pyramid network. The angle feature extraction module is used to generate multi-scale features through deep separable bar convolution after dimension reduction of the multi-scale features, and the multi-scale features are angle-selected and resampled by a rotation matrix to generate enhanced multi-scale features.

[0102] A third feature extraction module is configured to input the enhanced multi-scale features into the feature pyramid network, and output target multi-scale features in a five-layer pyramid structure through horizontal connection, layer-by-layer upsampling and convolution fusion operations.

[0103] A detection head setting module is configured to set a detection head on each multi-scale output of the feature pyramid network, which is used for classification, bounding box regression and angle parameter regression.

[0104] A detection module is configured to introduce a frequency domain channel attention mechanism in the angle regression branch through an angle prediction head module in the detection head. The target multi-scale features are mapped to the frequency domain through Fourier transform, restored to the spatial domain after channel attention, and the rotation angle is independently regressed based on integral operation.

[0105] An optimization module is configured to jointly optimize the regression information using ProbIoU loss and distributed focal loss to generate a detection result.

[0106] The angle attention-based anchor-free remote sensing image rotating target detection system provided by the embodiment extracts multi-scale features through a pre-trained backbone network and a feature pyramid network, strengthens the capture of target angle features by combining an angle feature extraction module and a frequency domain channel attention mechanism, realizes accurate prediction of target categories, positions, scales and rotation angles by means of an optimized detection head, and effectively improves the detection accuracy and efficiency of targets with various scales and directions in remote sensing images by jointly optimizing the regression information using ProbIoU loss and distributed focal loss, thereby enhancing the adaptability and robustness of the model in complex remote sensing scenarios.

[0107] Figure 5 A schematic diagram of a hardware structure of an electronic device for implementing various embodiments of the present application.

[0108] The angle attention-based anchor-free remote sensing image rotation target detection method provided by the embodiments of the present application can be applied to an electronic device. Those skilled in the art can understand that the electronic device structure involved in the embodiments of the present application does not constitute a limitation on the electronic device, and the electronic device can include more or fewer components than the diagram, or combine certain components, or different component arrangements. In the embodiments of the present application, the electronic device includes but is not limited to a laptop computer, a desktop computer, a workstation, a personal digital assistant, a server, a blade server, a mainframe computer, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the embodiments of the present application described and / or claimed herein.

[0109] The electronic device can include a processor, an external memory interface, an internal memory, a universal serial bus (USB) interface, a charge management module, a power management module, a battery, a wireless communication module, an audio module, a speaker, a microphone, a sensor module, a key, a camera, a display screen, and a SIM card interface, etc.

[0110] The processor can include one or more processing units, such as: the processor can include a central processing unit (CPU), an application processor (AP), a modem processor, a graphics processing unit (GPU), an image signal processor (ISP), a controller, a memory, a video codec, a digital signal processor (DSP), a baseband processor, and / or a neural-network processing unit (NPU), etc. Among them, different processing units can be independent devices, or can be integrated in one or more processors.

[0111] Among them, the processor can be the nerve center and command center of the electronic device. The controller can generate operation control signals according to instruction operation codes and timing signals to complete the control of fetching and executing instructions.

[0112] The processor can further include a memory that stores instructions and data. In some embodiments, the memory in the processor is a cache memory. The memory can hold instructions or data that the processor has recently used or is likely to use again. If the processor needs to use the instructions or data again, it can call them directly from the memory. This avoids repeated access and reduces the processor's latency, thus improving system efficiency.

[0113] The external memory interface can be used to connect an external memory card, such as a MicroSD card, to extend the storage capacity of the electronic device. The external memory card communicates with the processor through the external memory interface to implement data storage functions. For example, music, video, and other files are saved in the external memory card.

[0114] The internal memory can be used to store computer executable program code, which includes instructions. The processor executes various functional applications and data processing of the electronic device by running the instructions stored in the internal memory. The internal memory can include a program storage area and a data storage area. The internal memory can include a high-speed random access memory and can also include a non-volatile memory, such as at least one magnetic disk storage device, a flash memory device, a universal flash storage (UFS), etc.

[0115] The wireless communication function of the electronic device can be implemented through an antenna, a wireless communication module, a modem processor, and a baseband processor, etc.

[0116] The wireless communication module can provide a wireless communication solution including wireless local area networks (WLAN) (such as a wireless fidelity (Wi-Fi) network), Bluetooth (BT), a global navigation satellite system (GNSS), frequency modulation (FM), near field communication (NFC), infrared technology (IR), etc. applied to the electronic device.

[0117] The electronic device can implement audio functions, etc. through an audio module, a speaker, a receiver, a microphone, an earphone interface, and an application processor, etc.

[0118] The electronic device can implement a shooting function through an ISP, a camera, a video codec, a GPU, a display screen, and an application processor, etc.

[0119] The electronic device can realize display function through GPU, display screen and application processor.

[0120] The GPU is a microprocessor for image processing, connected to the display screen and the application processor. The GPU is used to perform mathematical and geometric calculations for graphics rendering. The processor can include one or more GPUs that execute program instructions to generate or change display information.

[0121] The display screen is used to display images, videos, etc. The display screen includes a display panel.

[0122] The electronic device realizes the anchor-free frame remote sensing image rotating target detection method based on angle attention provided in the application. By using the pre-trained backbone network and the feature pyramid network for multi-scale feature extraction, the angle feature extraction module is increased and the frequency domain channel attention mechanism is introduced to strengthen the angle feature processing, the optimized detection head is set for target multi-parameter prediction, and the ProbIoU loss and the distributed focal loss are used to jointly optimize the regression information, which improves the remote sensing image rotating target detection precision and efficiency, enhances the adaptability and robustness of the model to complex scenes.

[0123] In the storage medium provided in the application, a program product capable of realizing an anchor-free frame remote sensing image rotating target detection method based on angle attention is stored.

[0124] The anchor-free frame remote sensing image rotating target detection method based on angle attention comprises: inputting a remote sensing image and extracting multi-scale features using a pre-trained backbone network; adding an angle feature extraction module between the backbone network and the feature pyramid network, using the angle feature extraction module to reduce the dimension of the multi-scale features, and then generating multi-scale features in parallel through deep separable strip convolution after the dimension reduction, and combining a rotation matrix to select and resample the multi-scale features to generate enhanced multi-scale features; inputting the enhanced multi-scale features into the feature pyramid network, outputting target multi-scale features in a five-layer pyramid structure through horizontal connection, layer-by-layer upsampling and convolution fusion operation; respectively connecting detection heads on the multi-scale outputs of the feature pyramid network for classification, bounding box regression and angle parameter regression; in the detection head, through an angle prediction head module, introducing a frequency domain channel attention mechanism in the angle regression branch, mapping the target multi-scale features to the frequency domain through Fourier transform, restoring to the spatial domain after channel attention using inverse Fourier transform, and independently regressing the rotation angle based on integral operation; using ProbIoU loss and distributed focal loss to jointly optimize the regression information to generate detection results.

[0125] In some possible implementation manners, the angle attention based anchor-free remote sensing image rotation target detection method of the present disclosure can be implemented in the form of a program product, which includes program codes for causing an end device to perform the steps described in the above “Exemplary Method” section of the present specification according to various exemplary embodiments of the present disclosure when the program product is run on the end device.

[0126] The storage medium of the present disclosure can adopt any combination of one or more readable media. The readable medium can be a readable signal medium or a readable storage medium. The readable storage medium may, for example, be but is not limited to an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or apparatus, or any combination thereof. More specific examples (non-exhaustive list) of the readable storage medium include an electrical connection having one or more wires, a portable disc, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof.

[0127] The above description of disclosed embodiments enables one of ordinary skill in the art to make or use the application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein can be applied to other embodiments without departing from the spirit or scope of the application. Thus, the present application is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for angle attention based anchor-free remote sensing image rotation target detection, characterized in that, The method comprises the following steps: inputting a remote sensing image and extracting multi-scale features using a pre-trained backbone network; adding an angle feature extraction module between the backbone network and the feature pyramid network, using the angle feature extraction module to reduce the dimension of the multi-scale features, and then generating multi-scale features in parallel through a depth separable bar convolution, and combining a rotation matrix to select and resample the multi-scale features to generate enhanced multi-scale features; inputting the enhanced multi-scale features into the feature pyramid network, and outputting target multi-scale features in a five-layer pyramid structure through horizontal connection, layer-by-layer upsampling and convolution fusion operations; respectively connecting detection heads on the multi-scale outputs of the feature pyramid network for classification, bounding box regression and angle parameter regression; in the detection head, an angle prediction head module is used to introduce a frequency domain channel attention mechanism in the angle regression branch, map the target multi-scale features to the frequency domain through Fourier transform, restore them to the spatial domain after channel attention, and independently regress the rotation angle based on integral operation; using ProbIoU loss and distributed focal loss to jointly optimize the regression information to generate detection results.

2. The angle attention based anchor-free remote sensing image rotation target detection method according to claim 1, characterized in that, The method comprises the following steps: processing a remote sensing image inputting a pre-trained ResNet-50 backbone network; The ResNet-50 backbone network comprises convolutional layers, pooling layers, four-stage residual modules and bottleneck structures. The ResNet-50 backbone network extracts three-layer feature maps C3, C4 and C5 of different scales in the remote sensing image to generate multi-scale features.

3. The angle attention based anchor-free remote sensing image rotation target detection method according to claim 2, characterized in that, The method comprises the following steps: The input multi-scale feature feature extraction is performed; wherein H, W represent the number of channels, spatial height and width, respectively represents the first group of feature maps output by the ResNet-50 backbone network, with a value range of 3, 4, and 5; the input multi-scale feature is reduced in channel number by a 1x1 convolutional layer to obtain a compact representation ; Contextual information is captured by multi-branch convolutional structures and relationships between different channels are modeled; The output of the 1x1 convolution is directly used as attention weights to weight the input of the 1x1 convolution. The angle feature extraction process is represented as follows: wherein and are the attention map and output of feature extraction, respectively, is an element-wise matrix multiplication operation, and DW_Conv denotes a depthwise convolution, denotes the branch of the multi-branch convolution structure, and scale0 is a residual connection.

4. The angle attention based anchor-free remote sensing image rotation target detection method according to claim 3, characterized in that, The method comprises the following steps: The enhanced multi-scale features are denoted as and input into the feature pyramid network; To 1x1 convolutional dimensionality reduction, uniform channel number, top-down layer by layer up-sampling and element-wise addition with the next layer feature, and then smoothed by 3x3 convolution, finally get pyramid features P3, P4, P5; In the feature map of P5, 3x3 convolution is performed for layer-by-layer downsampling to obtain pyramid features P6 and P7, which can be formally represented as: wherein, represents the first layer FPN output feature map, is the output feature map of the corresponding layer in the angle feature extraction module, Upsample(·) represents an up-sampling operation, the value range of is 3, 4, 5; Based on the pyramid features P3, P4, P5, P6 and P7, target multi-scale features are generated.

5. The angle attention based anchor-free remote sensing image rotation target detection method according to claim 4, characterized in that, The method comprises the following steps: The pyramid features P3, P4, P5, P6 and P7 output by the feature pyramid network are respectively connected with detection heads for classification, bounding box regression and angle parameter regression. The detection head branches are respectively set on the pyramid features P3, P4, P5, P6 and P7 output by the feature pyramid network; the detection head branches comprise parallel classification branches, bounding box regression branches and direction branches; The classification branch is used to determine whether there is a target at each position and give the class probability; The bounding box regression branch is used to predict the center position coordinates (x, y) and scale size (w, h) parameters of the target bounding box; wherein w is the width of the target bounding box and h is the height of the target bounding box; 6. The angle attention based anchor-free remote sensing image rotation target detection method according to claim 5, characterized in that, The direction branch is used to predict the rotation angle parameter θ of the target. The frequency domain channel attention mechanism comprises: The channel attention is adjusted by the following formula: wherein, represents a sigmoid function, represents a fully connected layer, represents the obtained multispectral vector.

7. The angle attention based anchor-free remote sensing image rotation target detection method according to claim 6, characterized in that, The regression information is optimized by using the ProbIoU loss and the distributed focal loss, including: The regression information is optimized by using the ProbIoU loss, and the regression information includes target position (x, y), size (w, h), and rotation angle θ; the rotation angle θ is predicted by using the distributed focal loss; The ProbIoU loss has the form of: where H i represents the probability distribution value of the predicted bounding box at the pixel point i; G i represents the probability distribution value of the real bounding box at the pixel point i; is a smoothing term to prevent numerical instability; The distributed focal loss has the form of: wherein, denotes the probability that the angle decreases within each interval, is set to π / 180.

8. An angle attention based anchor-free remote sensing image rotating target detection system, characterized in that, The system adopts the anchor-free remote sensing image rotating target detection method based on angle attention according to any one of claims 1 to 7. The system comprises: The first feature extraction module is configured to input the remote sensing image and extract multi-scale features by using a pre-trained backbone network; The second feature extraction module is configured to add an angle feature extraction module between the backbone network and a feature pyramid network, generate multi-scale features by using the angle feature extraction module to reduce dimensions of the multi-scale features and then passing the reduced multi-scale features through deep separable convolution in parallel, and generate enhanced multi-scale features by using a rotation matrix to select angles and resample the multi-scale features; The third feature extraction module is configured to input the enhanced multi-scale features into the feature pyramid network, output target multi-scale features in a five-layer pyramid structure by using horizontal connection, layer-by-layer upsampling, and convolution fusion operations; The detection head setting module is configured to connect detection heads to multi-scale outputs of the feature pyramid network respectively, and perform classification, bounding box regression, and angle parameter regression; The detection module is configured to introduce a frequency domain channel attention mechanism in an angle regression branch by using an angle prediction head module in the detection head, map the target multi-scale features to a frequency domain by using Fourier transform, restore the target multi-scale features to a spatial domain by using inverse Fourier transform after channel attention, and independently regress the rotation angle based on integral operation. The optimization module is configured to optimize the regression information by using the ProbIoU loss and the distributed focal loss, and generate a detection result.

9. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor implements the steps of the anchor-free remote sensing image rotating target detection method based on angle attention according to any one of claims 1 to 7 when executing the program.

10. A storage medium having stored thereon a computer program, characterized in that The computer program implements the steps of the anchor-free remote sensing image rotating target detection method based on angle attention according to any one of claims 1 to 7 when executed by the processor.

Images