Closed-loop fused double-branch cooperative rotation remote sensing target detection network and method

Abstract

The invention provides a closed-loop fused double-branch cooperative rotation remote sensing target detection network and method, and relates to the technical field of image processing. The network comprises: a convolution branch configured as a multi-stage residual network, each stage of the residual network at least comprising a first bottleneck block and a second bottleneck block; the attention branch is configured to be a multi-stage Swin Transform, and each stage of the Swin Transform at least comprises a Swin block; a bidirectional multi-point fusion mechanism is arranged between the convolution branch and the attention branch and comprises a forward fusion mechanism from the convolution branch to the attention branch and a reverse fusion mechanism from the attention branch to the convolution branch; and the output of the network is subjected to multi-scale fusion through the feature pyramid network and then sent to a detection head to complete rotating target detection. According to the network, the convolution shallow layer edge and the attention deep layer context are optimized in the same field through reverse recharging, and the positioning and angle regression precision is improved.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically, to a closed-loop fusion dual-branch cooperative rotating remote sensing target detection network and method. Background Technology

[0002] Remote sensing images often present challenges such as complex sea and sky backgrounds, large scale spans, and difficulty in determining target location. Traditional detectors based on natural scenes suffer significant performance degradation in such scenarios, making them unsuitable for application requirements.

[0003] Existing hybrid architectures of Convolutional Neural Networks (CNNs) and attention-based deep learning models (such as Transformers) often employ unidirectional feature injection, introducing convolutional features at a few locations to enhance local information. However, this approach struggles to guarantee consistent feature distribution and semantic coherence between the two branches, and its versatility and substitutability are limited. Therefore, there is an urgent need for a unified network that jointly maintains both local sensitivity and global semantics to overcome the shortcomings of existing technologies. Summary of the Invention

[0004] In view of this, the present invention provides a closed-loop fusion dual-branch cooperative rotating remote sensing target detection network and method.

[0005] One aspect of the present invention provides a closed-loop fusion dual-branch cooperative rotating remote sensing target detection network, comprising: a convolutional branch configured as a multi-stage residual network, each stage of the residual network including at least a first bottleneck block and a second bottleneck block; and an attention branch configured as a multi-stage Swing Transformer, each stage of the Swing Transformer including at least one Swing block; wherein a bidirectional multi-point fusion mechanism is provided between the convolutional branch and the attention branch, the bidirectional multi-point fusion mechanism including: a forward fusion mechanism from the convolutional branch to the attention branch, and a reverse fusion mechanism from the attention branch to the convolutional branch; the forward fusion mechanism is configured to... Within the first Swing block of each stage of the Transformer: in response to the window multi-head self-attention computation of the Swing block, the convolutional output of the first bottleneck block in the convolutional branch of the corresponding stage is injected; and in response to the first layer normalization operation of the Swing block, the convolutional output of the second bottleneck block in the convolutional branch of the corresponding stage is injected; the inverse fusion mechanism is configured to fuse the output of the multilayer perceptron after the window multi-head self-attention computation of each Swing block and the output of the multilayer perceptron after the shift window multi-head self-attention computation to the output side and input side of the second bottleneck block of the corresponding stage of the convolutional branch, respectively.

[0006] According to an embodiment of the present invention, in the bidirectional multi-point fusion mechanism, the injection intensity is configured to be jointly controlled by channel-level gating and stage-level gating.

[0007] According to an embodiment of the invention, channel-level gating is configured to be generated by compression and excitation networks; stage-level gating is configured to be generated by learnable scalars for each stage and sparse regularization is applied to suppress redundant injection.

[0008] According to an embodiment of the present invention, the second bottleneck block represents the second bottleneck unit arranged sequentially in the residual network of each stage. The second bottleneck unit is configured as a 3×3 convolutional layer, and the injection point of the second bottleneck unit is fixed at the input side and the output side of the 3×3 convolutional layer.

[0009] According to an embodiment of the present invention, the output side of the network is connected to a general feature pyramid network, which is used to perform multi-scale fusion of the network output and then send it to the detection head for rotating target detection.

[0010] According to an embodiment of the present invention, the detection head is a rotating frame detection head, which is represented by five parameters, wherein the angle regression and classification branch share the output of the general feature pyramid network.

[0011] According to an embodiment of the present invention, the output channels of each stage of the convolution branch are 256 / 512 / 1024 / 2048, respectively, and the embedding dimensions projected onto the attention branch are 96 / 192 / 384 / 768, respectively.

[0012] According to an embodiment of the present invention, the network is configured with a cross-branch consistency loss function and an angle consistency loss function.

[0013] According to an embodiment of the present invention, the cross-branch consistency loss function adopts cosine similarity or distribution matching loss; the angle consistency loss function is used to impose consistency constraints on the classification angle and the regression angle.

[0014] Another aspect of the present invention provides a closed-loop fusion-based dual-branch collaborative rotational remote sensing target detection method, comprising: acquiring a remote sensing image and feeding it into a convolutional branch, extracting the output of the first bottleneck block and the output of the second bottleneck block at each stage respectively; performing 1×1 convolutional projection, two-level normalization, and pooling and position adaptive interpolation alignment processing on the first bottleneck block output and the second bottleneck block output, and injecting them into the first Swin block of the attention branch of the corresponding stage, respectively before the window multi-head self-attention calculation and after the first multilayer perceptron; acquiring the output of the multilayer perceptron after the window multi-head self-attention calculation of each Swin block, and the output of the multilayer perceptron after the shift window multi-head self-attention calculation, and after reshaping, interpolation, 1×1 convolutional projection, and batch normalization, and fusing them into the output side and input side of the second bottleneck block of the corresponding stage of the convolutional branch respectively; controlling the injection intensity through channel-level gating and stage-level gating; feeding the output features of each stage into a general feature pyramid network for multi-scale fusion, and then feeding them into a detection head for rotating target detection to obtain the category and position of the rotating target.

[0015] Compared with the prior art, the closed-loop fusion dual-branch cooperative rotating remote sensing target detection network, method, and apparatus provided in this embodiment of the invention have at least the following beneficial effects:

[0016] (1) The closed-loop fusion dual-branch collaborative rotational remote sensing target detection network and method provided in this embodiment of the invention, while ensuring structural stability, improves the collaborative expression of convolutional local details and Transformer global semantics in complex rotation and large-scale change scenarios through closed-loop bidirectional information interaction and fixed injection source implementation rules, avoids distribution shift and training instability caused by unidirectional injection, enhances feature extraction of the target during remote sensing target detection, thereby improving the accuracy of remote sensing image target detection.

[0017] (2) The closed-loop fusion dual-branch collaborative rotational remote sensing target detection network and method provided in the embodiments of the present invention adopts an artificial intelligence platform, learnable alignment, and channel / stage dual gating, which reduces the distribution shift caused by cross-branch and makes training more stable.

[0018] (3) The closed-loop fusion dual-branch collaborative rotating remote sensing target detection network and method provided in the embodiments of the present invention can be directly connected to the general feature pyramid network, and has high engineering substitutability and deployment friendliness.

[0019] (4) The closed-loop fusion dual-branch cooperative rotating remote sensing target detection network and method provided in the embodiments of the present invention has high generalization and robustness, and performs stably under complex backgrounds, scale spans and arbitrary directions. It can be widely used for rotating target detection such as ships, aircraft and vehicles. Attached Figure Description

[0020] The above and other objects, features and advantages of the present invention will become more apparent from the following description of embodiments of the invention with reference to the accompanying drawings, in which:

[0021] Figure 1 This diagram schematically illustrates the structure of a closed-loop fusion dual-branch cooperative rotating remote sensing target detection network according to an embodiment of the present invention.

[0022] Figure 2 A schematic diagram of a bidirectional gated fusion injection module within a single stage according to an embodiment of the present invention is shown.

[0023] Figure 3 A flowchart illustrating a closed-loop fusion dual-branch cooperative rotating remote sensing target detection method according to an embodiment of the present invention is shown. Detailed Implementation

[0024] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the invention. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of the invention for ease of explanation. However, it will be apparent that one or more embodiments may be practiced without these specific details. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0025] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0026] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0027] When using expressions such as "at least one of A, B and C", they should generally be interpreted in accordance with the meaning that is commonly understood by those skilled in the art (e.g., "a system having at least one of A, B and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B and C, etc.).

[0028] In the embodiments of this invention, the collection, updating, analysis, processing, use, transmission, provision, disclosure, and storage of data (e.g., including but not limited to user personal information) comply with relevant laws and regulations, are used for legitimate purposes, and do not violate public order and good morals. In particular, necessary measures have been taken to prevent unauthorized access to user personal information data and to maintain the security of user personal information and network security.

[0029] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0030] Figure 1 The diagram illustrates the structure of a closed-loop fusion dual-branch cooperative rotating remote sensing target detection network according to an embodiment of the present invention.

[0031] like Figure 1 As shown, the structure of the closed-loop fusion dual-branch cooperative rotating remote sensing target detection network in this embodiment of the invention includes: a convolutional branch and an attention branch.

[0032] The convolutional branch is configured as a four-stage residual network, such as a deep convolutional neural network (Residual Network 50-layer, ResNet50), each stage of which includes at least a first bottleneck block and a second bottleneck block.

[0033] The attention branch is configured as a four-stage Swin Transformer (a hierarchical visual model), each stage of which includes at least one Swin block, such as Block1, Block2, Block3, and Block4.

[0034] In this embodiment, a bidirectional four-point fusion mechanism is set between the convolution branch and the attention branch. The bidirectional four-point fusion mechanism includes a forward fusion mechanism from the convolution branch to the attention branch and a reverse fusion mechanism from the attention branch to the convolution branch.

[0035] This forward fusion mechanism (denoted as A) is configured to perform the following operations within the first Swin block of each stage of the Swin Transformer:

[0036] (A1) In response to the window multi-head self-attention (W-MSA) computation of the Swing block, inject the convolution output of the first bottleneck block in the corresponding stage convolution branch.

[0037] For example, before W-MSA computation, the 3×3 convolution (conv2) output of the first bottleneck block of the corresponding stage is preprocessed (such as 1×1 convolution projection, two-level normalization, and pooling and position adaptive interpolation alignment) and then injected.

[0038] (A2) In response to the first layer normalization (LN) operation of the Swin block, the convolution output of the second bottleneck block in the corresponding stage convolution branch is injected.

[0039] For example, after the first LN operation of the Swin block, the 3×3 convolution (conv2) output of the second bottleneck block of the corresponding stage is injected in the same way as the preprocessing of (A1). That is, after the first LN operation of the Swin block, the convolution output of the second bottleneck block in the convolution branch of the corresponding stage is injected after 1×1 convolution projection, two-level normalization, pooling and position adaptive interpolation alignment.

[0040] The reverse fusion mechanism (denoted as B) is configured to combine each Swin block as follows:

[0041] The outputs of the multi-layer perceptron (MLP) after (B1) multi-head self-attention calculation and (B2) multi-layer perceptron after (B2) shifted window multi-head self-attention (SW-MSA) calculation are preprocessed and then fused to the output side (back side) and input side (front side) of the second bottleneck block in the corresponding stage of the convolution branch, respectively.

[0042] For example, the (B1) W-MSA MLP output and (B2) SW-MSA MLP output of each Swing block are preprocessed (such as sequence image reshaping, size interpolation, 1×1 convolution and batch normalization) and then fused to the back and front sides of the second bottleneck block conv2 of the corresponding stage of the convolution branch.

[0043] The output of the entire network is fused at multiple scales by Feature Pyramid Networks (FPN) and then fed into the detection head to complete the detection of rotating targets.

[0044] The closed-loop fusion dual-branch collaborative rotational remote sensing target detection network provided in this embodiment of the invention, while ensuring structural stability, improves the collaborative expression of convolutional local details and Transformer global semantics in complex rotation and large-scale change scenarios through closed-loop bidirectional information interaction and fixed injection source implementation rules. It avoids distribution shift and training instability caused by unidirectional injection, enhances feature extraction of the target during remote sensing target detection, and thus improves the accuracy of target detection in remote sensing images.

[0045] According to an embodiment of the present invention, in the bidirectional multi-point fusion mechanism, the injection intensity is configured to be gated by the channel level (α). c ) and staged gating (α) s Joint control.

[0046] Among them, the channel-level gating (α) c The ) is configured to be generated via a squeeze-and-excitation network, for example, it can be generated by global average pooling with two fully connected layers.

[0047] This stage-level gating (α) s It is configured to be generated via a learnable scalar at each stage and sparse regularization is applied to suppress redundant injection, for example, stage-level gating (α). s It can perform channel-level gating (α) c Apply L1 regularization.

[0048] In this embodiment, channel-level gating (α) c ) and staged gating (α) s The functionality can be implemented based on a bidirectional gated fusion injection module, the schematic diagram of which can be found in [reference needed]. Figure 2 As shown.

[0049] Figure 2 A schematic diagram of a bidirectional gated fusion injection module within a single stage according to an embodiment of the present invention is shown.

[0050] like Figure 2 As shown, the left side displays a stage (stage input) of a ResNet50 network, containing two bottleneck blocks (block1 and block2). The right side displays a stage (stage input) of a Swin Transformer, containing two Swin blocks.

[0051] Regarding the forward fusion of convolutional branches to attention branches:

[0052] Information source: Outputs from ResNet50 network block 1 and block 2.

[0053] Processing flow:

[0054] (1) Downsampling and channel projection.

[0055] (2) Average pooling, further adjusting the space size.

[0056] (3) Reshape the 2D feature map into a 1D sequence to fit the input format of the Transformer.

[0057] (4) Layer normalization.

[0058] Injection point: The processed CNN features are injected before the W-MSA / SW-MSA of the Swing Transformer block.

[0059] Regarding the reverse fusion of attention branches to convolution branches:

[0060] Information source: MLP output from two Swing blocks of Swing Transformer.

[0061] Processing flow:

[0062] (1) Reconstruct the 1D sequence of the Transformer back into a 2D feature map.

[0063] (2) Interpolation upsampling: Adjust the spatial resolution of the Transformer features to be the same as the feature map of the corresponding stage of the CNN.

[0064] (3) Channel projection (adjusting the embedding dimension of the Transformer to the number of channels required inside the bottleneck block of the CNN).

[0065] (4) Batch standardization.

[0066] Injection points: The processed Transformer features are injected into the "front and back sides" of the conv:3*3 convolutional layer (i.e., conv2) of ResNet block1 and block2.

[0067] Front side: Injected into the conv:3*3 layer as part of its input.

[0068] Back side: Output injected into the conv:3*3 layer.

[0069] The closed-loop fusion dual-branch collaborative rotating remote sensing target detection network provided in this embodiment of the invention adopts an artificial intelligence platform, learnable alignment, and channel / stage dual gating, which reduces the distribution shift caused by cross-branch, making training more stable.

[0070] According to an embodiment of the present invention, the second bottleneck block represents the second bottleneck unit arranged sequentially in the residual network of each stage. The second bottleneck unit conv2 is configured as a 3×3 convolutional layer, and the injection point of the second bottleneck unit is fixed at the input side (front side) and output side (back side) of the 3×3 convolutional layer.

[0071] According to an embodiment of the present invention, the feature pyramid network can be a general feature pyramid network, and the output side of the entire network (backbone network) is connected to the general feature pyramid network.

[0072] This general feature pyramid network is used to fuse the network's output at multiple scales before feeding it into the detection head for rotating target detection.

[0073] In this embodiment, the general feature pyramid network can be configured with four levels, such as P2, P3, P4, and P5. The backbone network also has four output levels, such as C2, C3, C4, and C5.

[0074] After the backbone network outputs four layers of features, the general feature pyramid network first performs 1x1 convolutions on the four layers of features (C2, C3, C4, C5) output by the backbone network, mainly to unify the number of channels. Then, starting from the deepest layer (C5), it progressively moves upwards, upsampling the deep features and adding them element-wise with the shallow features of the corresponding layers. After each fusion, a 3x3 convolution is used for smoothing, generating the final P5, P4, P3, P2 pyramid features.

[0075] The closed-loop fusion dual-branch collaborative rotating remote sensing target detection network provided in this embodiment of the invention has its output directly connected to a general feature pyramid network, and has high engineering replaceability and deployment friendliness.

[0076] According to an embodiment of the present invention, the detection head is a rotating frame detection head, which is represented by five parameters, wherein the angle regression and classification branch share the output of the general feature pyramid network.

[0077] According to an embodiment of the present invention, the output channels of each stage of the convolution branch are 256 / 512 / 1024 / 2048, respectively, and the embedding dimensions projected onto the attention branch are 96 / 192 / 384 / 768, respectively.

[0078] In this embodiment, the output channels of each stage of the convolution branch are as follows:

[0079] The feature map output by Stage 2 has 256 channels.

[0080] The feature map output by Stage 3 has 512 channels.

[0081] The feature map output by Stage 4 has 1024 channels.

[0082] The feature map output by Stage 5 has 2048 channels.

[0083] The output of each stage of the convolutional branch, when injected into the corresponding stage of the attention branch, needs to have its channel count adjusted to match the embedding dimension of that stage of the attention branch. For example:

[0084] The features of the convolution branch Stage 2 (256 channels) are projected onto 96 dimensions when injected into the attention branch Stage 2.

[0085] The features of the convolution branch Stage 3 (512 channels) are projected to 192 dimensions when injected into the attention branch Stage 3.

[0086] The features of the convolution branch Stage 4 (1024 channels) are projected to 384 dimensions when injected into the attention branch Stage 4.

[0087] The features of the convolution branch Stage 5 (2048 channels) are projected onto the 768-dimensional scale when injected into the attention branch Stage 5.

[0088] According to an embodiment of the present invention, the overall network is configured with a cross-branch consistency loss function L_align and an angle consistency loss function L_theta.

[0089] The cross-branch consistency loss function uses cosine similarity or distribution matching loss; the angle consistency loss function is used to impose consistency constraints on the classification angle and the regression angle.

[0090] The closed-loop fusion dual-branch cooperative rotating remote sensing target detection network provided in this embodiment of the invention has high generalization and robustness, and performs stably under complex backgrounds, scale spans and arbitrary directions. It can be widely used for the detection of rotating targets such as ships, aircraft and vehicles.

[0091] Based on the above closed-loop fusion dual-branch cooperative rotating remote sensing target detection network, this embodiment of the invention also provides a closed-loop fusion dual-branch cooperative rotating remote sensing target detection method.

[0092] Figure 3 A flowchart illustrating a closed-loop fusion dual-branch cooperative rotating remote sensing target detection method according to an embodiment of the present invention is shown.

[0093] like Figure 3 As shown, the closed-loop fusion dual-branch cooperative rotating remote sensing target detection method of this invention may include, for example, operations S1 to S5.

[0094] In operation S1, remote sensing images are acquired and fed into the convolution branch, and the first bottleneck block output and the second bottleneck block output of each stage are extracted respectively.

[0095] In operation S2, after performing 1×1 convolution projection, two-level normalization, pooling, and position adaptive interpolation alignment on the outputs of the first and second bottleneck blocks, the outputs are injected into the first Swing block of the corresponding stage attention branch, before the window multi-head self-attention calculation and after the first multilayer perceptron, respectively.

[0096] In operation S3, the output of the multilayer perceptron after window multi-head self-attention computation of each Swing block and the output of the multilayer perceptron after shift window multi-head self-attention computation are obtained. After reshaping, interpolation, 1×1 convolution projection, and batch normalization, they are respectively fused to the output side and input side of the second bottleneck block of the corresponding stage of the convolution branch.

[0097] In operation S4, the injection intensity is controlled by channel-level gating and stage-level gating.

[0098] In operation S5, the output features of each stage are fed into a general feature pyramid network for multi-scale fusion, and then fed into the detection head for rotating target detection to obtain the category and position of the rotating target.

[0099] In this embodiment, the original remote sensing image is first read in and fed into the convolutional branch of the bidirectional four-point fusion backbone network to obtain four-stage features. The outputs of the first bottleneck block conv2 and the second bottleneck block conv2 of each stage are then taken.

[0100] Then, the two types of features in S1 are projected, normalized, pooled, and positionally adaptively interpolated, and then injected into the first SwinBlock of the corresponding Swing stage before W-MSA and after the first MLP, respectively.

[0101] Then, the MLP outputs after W-MSA and SW-MSA of each SwinBlock are obtained, and after reshaping, interpolation, 1×1 convolution, and batch normalization, they are fused to the back and front sides of the second bottleneck block conv2 in the corresponding stage of the convolution branch.

[0102] With channel-level gating (α) c ) and staged gating (α) s Control the injection intensity.

[0103] Finally, the outputs from each stage are sent to the general FPN and then into the detection head, which outputs the category and position of the rotating target, thus completing the rotating target detection.

[0104] In this embodiment, the location-adaptive interpolation involves: predicting the sampling offset and weighting coefficients through lightweight convolution, and aligning them to the window resolution of the current Swin stage.

[0105] For example, location-adaptive interpolation can be aligned with the weights w by predicting the offset (Δu, Δv), with an interpolation kernel size of 3×3 or 5×5.

[0106] In this embodiment, the alignment pipeline for forward and reverse fusion includes "sequence-image reshaping-size protection (minimum 4×4)-interpolation-1×1 convolution-normalization", and triggers upsampling protection when the size is below a threshold.

[0107] In this embodiment, the input to the general feature pyramid network is the unified feature map after the bidirectional fusion operation of S2 and S3 has been completed.

[0108] The embodiments of the present invention have been described above. However, these embodiments are merely illustrative and not intended to limit the scope of the invention. Although various embodiments have been described above, this does not mean that the measures in the various embodiments cannot be used advantageously in combination. Various substitutions and modifications can be made by those skilled in the art without departing from the scope of the invention, and all such substitutions and modifications should fall within the scope of the invention.

Claims

1. A closed-loop fusion dual-branch cooperative rotating remote sensing target detection network, characterized in that, The network includes: The convolutional branch is configured as a multi-stage residual network, each stage of which includes at least a first bottleneck block and a second bottleneck block. The attention branch is configured as a multi-stage Swing Transformer, each stage of which includes at least one Swing block; The convolutional branch and the attention branch are provided with a bidirectional multi-point fusion mechanism, which includes a forward fusion mechanism from the convolutional branch to the attention branch and a reverse fusion mechanism from the attention branch to the convolutional branch. The forward fusion mechanism is configured to occur within the first Swing block of each stage of the Swing Transformer: In response to the multi-head self-attention computation of the window in the Swing block, the convolution output of the first bottleneck block in the corresponding stage convolution branch is injected; and In response to the first layer normalization operation of the Swin block, the convolution output of the second bottleneck block in the corresponding stage convolution branch is injected; The reverse fusion mechanism is configured to fuse the output of the multilayer perceptron after window multi-head self-attention calculation of each Swing block and the output of the multilayer perceptron after shift window multi-head self-attention calculation to the output side and input side of the second bottleneck block of the corresponding stage of the convolution branch, respectively.

2. The network according to claim 1, characterized in that, In the bidirectional multi-point fusion mechanism, the injection intensity is configured to be jointly controlled by channel-level gating and stage-level gating.

3. The network according to claim 2, characterized in that, The channel-level gating is configured to be generated via a compression and excitation network; the stage-level gating is configured to be generated via a learnable scalar at each stage and to apply sparse regularization to suppress redundant injection.

4. The network according to claim 1, characterized in that, The second bottleneck block represents the second bottleneck unit arranged sequentially in the residual network of each stage. The second bottleneck unit is configured as a 3×3 convolutional layer, and the injection point of the second bottleneck unit is fixed at the input side and the output side of the 3×3 convolutional layer.

5. The network according to claim 1, characterized in that, The output side of the network is connected to a general feature pyramid network, which is used to perform multi-scale fusion of the network's output and then send it to the detection head for rotating target detection.

6. The network according to claim 5, characterized in that, The detection head is a rotating frame detection head, represented by five parameters, wherein the angle regression and classification branches share the output of the general feature pyramid network.

7. The network according to claim 1, characterized in that, The output channels of each stage of the convolution branch are 256 / 512 / 1024 / 2048, and the corresponding embedding dimensions projected onto the attention branch are 96 / 192 / 384 / 768.

8. The network according to claim 1, characterized in that, The network is configured with a cross-branch consistency loss function and an angle consistency loss function.

9. The network according to claim 8, characterized in that, The cross-branch consistency loss function uses cosine similarity or distribution matching loss; the angle consistency loss function is used to impose consistency constraints on the classification angle and the regression angle.

10. A closed-loop fusion dual-branch cooperative rotating remote sensing target detection method, applied to the closed-loop fusion dual-branch cooperative rotating remote sensing target detection network as described in any one of claims 1 to 9, characterized in that, The method includes: The remote sensing image is acquired and fed into the convolution branch, and the first bottleneck block output and the second bottleneck block output of each stage are extracted respectively. After performing 1×1 convolution projection, two-level normalization, pooling and position adaptive interpolation alignment on the outputs of the first bottleneck block and the outputs of the second bottleneck block, they are injected into the first Swing block of the corresponding stage attention branch, respectively before the window multi-head self-attention calculation and after the first multilayer perceptron. The output of the multilayer perceptron after window multi-head self-attention computation for each Swing block and the output of the multilayer perceptron after shift window multi-head self-attention computation are obtained, and after reshaping, interpolation, 1×1 convolutional projection, and batch normalization, they are fused to the output and input sides of the second bottleneck block of the corresponding stage of the convolutional branch, respectively. Injection intensity is controlled through channel-level gating and stage-level gating; The output features from each stage are fed into a general feature pyramid network for multi-scale fusion, and then fed into a detection head for rotating target detection to obtain the category and position of the rotating target.

Images