A remote sensing rotation detection method based on context feature adaptive weighted fusion

Abstract

The application discloses a remote sensing rotation detection network based on context feature adaptive weighted fusion. Mainly includes: (1) a one-dimensional Gaussian angle label encoder (GALE) is designed to encode the angle in the OBB label, and an angle classification loss (ACL) is proposed; (2) GALE, ACL and YOLOv5 are combined to construct a rotation target detector baseline GLR-YOLO; (3) a double-branch context adaptive modeling module (DBCAM) is designed to model, convert and fuse the context features, so as to reconstruct the feature extraction network; (4) an adaptive global feature aggregation and reweighting module (AGFAR) is proposed, and an adaptive detection head (ADH) is constructed in combination with the detection head to solve the inconsistency problem between output layers. The application is mainly used in the field of remote sensing image target detection and has important research value in the fields of building early warning, traffic monitoring and national defense safety.

Description

TECHNICAL FIELD

[0001] The application relates to a remote sensing rotating detection network based on context feature adaptive weighted fusion, in particular to a target detection problem in the field of remote sensing images, and belongs to the field of computer vision and intelligent information processing. BACKGROUND

[0002] Target detection is a challenging task in computer vision, aiming to identify and classify targets in images. In the past decade, target detection methods based on deep learning and convolutional neural networks have been continuously innovated. As a more specialized task in remote sensing image (RSI) detection, the rotating object detection (ROD) task has increasingly attracted the attention of researchers.

[0003] RSI refers to obtaining information about objects without contact at a very long distance through different imaging sensors, and the acquired visible light images have the characteristics of high resolution, clear targets, and optional observation range. Remote sensing ROD provides a new detection method for the target detection field and has important research value in the fields of building early warning, traffic monitoring, military and national defense security. Unlike traditional target detection techniques that mainly use horizontal bounding boxes (HBB), in RSI, due to its wide coverage, multiple target categories in the image, uneven distribution of different categories, and a large amount of noise carried by the image background, imaging is complex and targets are easily isolated, so target detection algorithms in natural scenes are difficult to achieve good results in remote sensing images. In addition, the overlap between two HBBs in RSI can be very large, making non-maximum suppression (NMS) based on HBB unsuitable for processing oriented objects. Therefore, the research on ROD methods using oriented bounding boxes (OBB) is necessary for accurate target detection in RSI. In view of the above problems, the application takes the target detection task in the field of RSI as the research object and proposes a remote sensing rotating detection network based on context feature adaptive weighted fusion, aiming to ensure the detection speed while improving the detection accuracy and effect of small targets in RSI. SUMMARY

[0004] The application provides a remote sensing rotating detection network (DAG-YOLO) based on context feature adaptive weighted fusion, which mainly comprises a one-dimensional Gaussian angle label encoder (1D Gaussian Angle Label-Encoder, GALE), a dual branch context adaptive modeling module (DBCAM) and an adaptive global feature aggregation and reweighting module (AGFAR). Specifically, the GALE is mainly responsible for encoding the angle label in the OBB label into 180 result values using a 1D Gaussian function, wherein the index of the angle label value is 1, and the remaining values conform to the 1D Gaussian distribution; the DBCAM uses channel self-attention and spatial self-attention to perform context adaptive modeling on the network input feature layer, fully captures the dual dependence relationship of the feature layer channel and space, improves the utilization of effective features, and reconstructs the feature extraction network; the AGFAR combines the network detection head to build an adaptive detection head, aggregates all output feature layers and reweights and fuses them, solves the inconsistency problem between different output layers, and realizes accurate prediction of large, medium and small targets. The cooperation of the above three parts enables the application to construct an efficient and accurate rotating target detection network, and realizes accurate positioning and identification of targets in the RSI.

[0005] A remote sensing rotating detection network based on context feature adaptive weighted fusion comprises the following steps:

[0006] (1) a one-dimensional Gaussian angle label encoder (GALE) is used to encode the angle in the OBB label into 180 results and all are reserved as the category label of the angle (containing 180 categories), and an angle classification loss (Angle Classification Loss, ACL) is designed;

[0007] (2) the one-dimensional Gaussian angle label encoder in (1) is combined with the YOLOv5 detection network to construct a rotating detector baseline GLR-YOLO, and the angle classification loss (Angle Classification Loss, ACL) is incorporated into the total loss of YOLOv5 to constrain and optimize the learning of GLR-YOLO parameters;

[0008] (3) In order to solve the problem that the feature is difficult to extract due to the complex target distribution in the remote sensing image (RSI), a double-branch context adaptive modeling module (DBCAM) is designed, a novel self-attention (channel + space) pooling context modeling method (SPCM) is proposed to obtain the feature space and channel dependence, and a double path weighted fusion method (DPWF) is designed to fully capture the dual dependence relationship between the channel and the space;

[0009] (4) In order to solve the problem that the features of three output layers of the detection network are inconsistent, an adaptive global feature aggregation and reweighting module (AGFAR) is designed, for each layer output, the feature shape adjustment module (FSA) is designed to adjust the size of the other two layers to the size of the layer, then the feature aggregation and reweighting module (FAR) is used to extract and reweight the weight feature map of the three feature layers, and the calculated weight is multiplied by each output feature layer to complete the feature reweighting, and finally the reweighted feature layers are added to complete the feature aggregation.

[0010] Compared with the prior art, the beneficial effects of the present application are:

[0011] 1、The present application adopts GALE, which converts the angle regression task into a classification task as a premise, obtains a more reasonable and accurate angle encoding representation, solves the boundary problem existing in the rotation detector, not only improves the performance of the detector, but also greatly reduces the optimization difficulty of the model;

[0012] 2、The ACL designed by the present application not only ensures the accuracy of angle prediction, but also balances the total loss of the rotation detector, enhances the robustness and stability of the model;

[0013] 3、The present application adopts DBCAM to reconstruct the feature extraction network, and uses self-attention pooling to model the context and capture and fuse the features, which controls the complexity of the model, reconstructs the feature extraction network, and enhances the network context feature extraction capability and the positioning capability of the target under high threshold;

[0014] 4、The present application designs an adaptive detection head combined with AGFAR, which reweights and fuses the most beneficial semantic information belonging to the same position of each feature layer through the relationship between feature mapping and scale transformation;

[0015] 5. The DAG-YOLO rotation detector constructed using a one-stage network in this invention not only maintains excellent speed performance but also achieves a stable improvement in detection performance, demonstrating its advantages and scalability in rotation detection tasks. Attached Figure Description

[0016] Figure 1 This is a block diagram of the overall model of the rotating target detection method of the present invention;

[0017] Figure 2 This is a structural diagram of the dual-branch context adaptive modeling module of the present invention;

[0018] Figure 3 This is a structural diagram of the adaptive global feature aggregation and reweighting module of the present invention;

[0019] Figure 4 This is a structural diagram of the feature shape alignment module of the present invention;

[0020] Figure 5 This is a structural diagram of the weight feature extraction module of the present invention;

[0021] Figure 6 This is a diagram showing the detection effect of the present invention in remote sensing images. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0023] like Figure 1 As shown, a rotating target detection method for RSI mainly consists of three parts: Backbone, Neck, and Head. The proposed DBCAM is fused with the C3 structure to form a C3D structure, which is then used to reconstruct the Backbone and Neck for feature extraction. The proposed AGFAR, combined with the Head, constructs an Adaptive Detection Head (ADH), solving the problem of inconsistent features between output layers. The network's final predicted scores for 180 angle categories and the GALE encoding results are used to calculate the loss using ACL, and other losses are combined to optimize the network parameters. Further explanations of each part of this invention follow:

[0024] 1. Rotating detector baseline GLR-YOLO

[0025] GLR-YOLO is a rotating detector based on YOLOv5, including the core improvement innovation points, namely GALE and ACL proposed by us. The idea of GALE is to convert the angle regression task into an angle classification task. The 180 channel values predicted by the network are used to fit the result value of the angle label encoding in OBB. The specific process is as follows:

[0026] Firstly, the original label (x1, y1, x2, y2, x3, y3, x4, y4) of the target marked by any quadrilateral in RSI is converted into OBB label (x, y, w, h, θ). In this application, OpenCV representation (D oc ) and Long-Edge representation (D le ) are used. Then, GALE is applied to encode the angle θ le in D le . The specific encoding process is given by Algorithm 1:

[0027]

[0028]

[0029] After obtaining the new label information (Rbox and GL), the loss is calculated using the new label and the positive samples extracted from the network prediction results, and then the gradient is back propagated to update the model parameters. In addition to the loss in YOLOv5, we additionally propose and introduce ACL. ACL depends on the BCEWithLogitsLoss function, which is represented as:

[0030]

[0031] Where n represents the matched positive sample; C represents that each sample has C categories; y c represents the GT label (0 or 1) of category c in sample n; x c represents the probability of the network model predicting that sample n belongs to category c.

[0032] The formula for obtaining the angle classification loss L θ using ACL is:

[0033]

[0034] Where θ P represents the 180 channel values predicted by the network model for each sample; θ GL represents the one-dimensional Gaussian encoding result of the angle in GT.

[0035] 2. Double-branch context adaptive modeling module (DBCAM)

[0036] Due to the complex target distribution in RSI, feature extraction and the utilization of cross-feature-layer context information bring challenges to the network. As shown in Figure 2 , we propose a dual-branch context adaptive modeling module (DBCAM) which consists of two parallel branches for context modeling of input features and feature transformation and fusion, respectively, S2C (the upper branch in DBCAM) and C2S (the lower branch in DBCAM), to fully capture the dual dependency between spatial dependency and channel dependency.

[0037] Since the two-branch structure performs simultaneously from input to output, we abstract the proposed method into two steps: first, adaptive feature context modeling; second, feature transformation and fusion. All the dimension changes of feature layers are shown in Figure 6 .

[0038] The adaptive feature context modeling process of S2C: the feature layer obtained after the image is extracted through multiple convolution layers is abstracted into C feature maps with the size of HxW, and a feature map is defined as where P = HW represents the number of positions on the feature map, and i represents the position index. Further, the entire input layer is defined as X:

[0039]

[0040] where x gi represents the position index of the feature information in the g feature map. The final output is defined as Z S2C . Therefore, the modeling formula of the spatial self-attention pool to obtain the context feature is defined as:

[0041]

[0042] where j, m represent the enumeration of all possible positions in a feature map; W k1 represents a linear transformation matrix; W k1 X represents the initial global feature weight. By calculating the feature mapping and shared weight, the global context feature modeling formula of the entire feature layer can be obtained:

[0043]

[0044] where, represents the modeling result of the context feature of a feature map.

[0045] The adaptive feature context modeling process of C2S: unlike the modeling process of S2C, we consider the feature vector from the channel dimension, redefine the input feature process as Y, and abstract it into P feature maps with the size of Cx1, so that the number of positions on a feature map is C. Therefore, we can define the feature mapping as where t denotes the position index. Further, Y is denoted as:

[0046]

[0047] where y ht denotes the feature information on the h-th feature map with position index t, and the output is defined as Z C2S Finally, we use the channel self-attention pooling to model the context feature, denoted as:

[0048]

[0049] where f, n denote the enumeration of all possible positions in a feature map; GAP, GMP represent global average pooling and global maximum pooling, respectively. β f denotes the global shared channel attention weight; W k2 Y denotes the initialized weight. The context modeling of an abstract feature map is extended to multiple feature maps, and the global feature context modeling formula is obtained:

[0050]

[0051] where h is from 1 to P, indicating that the context feature modeling is performed on each feature map of the abstract feature layer Y.

[0052] Feature conversion and fusion part: whether it is the S2C branch or the C2S branch, we use a double-path feature conversion module to correspond to two fusion weighting modes. Through feature transformation and double-path weighted fusion, S2C and C2S capture enhanced context channel dependency and context spatial dependency, respectively.

[0053] In the S2C branch, C / r (r=4) is the number of hidden channels. In order to reduce the optimization difficulty and improve the generalization ability, we use layer normalization (LN) before the ReLU function to fully capture the feature dependency between channels. In order to ensure the consistency of the distribution between features, we first use the Sigmoid function to connect the paths, and then use element-wise multiplication (EM) to recalibrate the original feature layer. Then, the aggregated features on other paths of this branch are weighted and fused into the recalibrated features in the form of broadcast element-wise addition (BEA). After global feature context modeling, feature conversion and fusion, the output of this branch is:

[0054] Z S2C = X Sigmoid (W v12 (ReLU (LN (W v11(ReLU(LN(W v22 (ReLU(LN(W v21 (ReLU(LN(W

[0055] where W v11 ,W v12 ,W 21 ,W 22 are convolutions with kernel size 1x1.

[0056] In the proposed C2S branch, we first use a large kernel convolution with kernel size 7x7, then use a ReLU function and a convolution with kernel size 3x3 to fully capture the spatial dependency of the features. In order to maintain the consistency of the feature distribution, the same double-path feature fusion method as S2C is adopted. Through global feature context modeling, feature transformation and weighted fusion, we define the C2S branch output as:

[0057] Z C2S = YU S1 = Sigmoid(U s12 (ReLU(U s11 (Ψ)))) + U s22 (ReLU(U s21 (Ψ))

[0058] where U s11 ,U s21 denote convolutions with kernel size 7x7, U s12 ,U s22 denote convolutions with kernel size 3x3.

[0059] The fusion results of the two branches are added to obtain the final output of DBCAM:

[0060] Z = Z S2C + Z C2S

[0061] Through global context modeling, fully capture the channel dependency and spatial dependency, combined with the feature fusion weighted two branch structure (S2C and C2S), provide useful semantic information supplement for the entire input feature.

[0062] 3. Adaptive global feature aggregation and re-weighting module (AGFAR)

[0063] In the network prediction process, the same position between different scale feature layers may carry contradictory information, which ultimately leads to inconsistent learning goals between different scale features. For example, Figure 3To solve the problem of feature inconsistency between FPN output layers and learn the feature information between different scale feature layers sufficiently, we propose AGFAR, which mainly consists of two parts: Feature Shape Alignment module (FSA) and Feature Aggregation Reweighting Module (FARM). To obtain the final aggregation feature layer (AL) of any scale, we define the input feature layer of this scale (e.g., P5) as the Dominant Layer of polymerization features (DL), and the other two input feature layers (P4 and P3) as Supplementary Layer of polymerization features (SLs).

[0064] As shown in Figure 4 , the role of FSA in AGFAR is to adjust the size of the three output layers to ensure that their feature mapping positions are consistent, so as to obtain the most useful information from the same position. The process is as follows: first, redefine the three output layers (P5, P4, P3) as (X L , X M , X S ). In each processing process, due to the different scales of the output layers as DL, although the method of feature aggregation reweighting is the same, the method of adjusting the feature shape of the two SLs to the shape of the DL is different, as shown in Figure 4 (a), (b), and (c) represent the alignment scheme of feature shape and DL, respectively. This scheme uses convolution layers (Conv) to adjust the number of channels, max-pooling layers (MP) and linear interpolation up-sampling layers (Upsample, UP) to adjust the size of the feature map. Each feature layer after FSA processing is defined as

[0065] The FARM module is the core step to solve the problem of feature inconsistency between output layers, and has two important components: Weight Feature Extraction (WFE) and Feature Aggregation Reweighting (FAR). The size of the feature layer as DL in the three output layers of the image is defined as CxHxW. After the FSA module, all three output layers are adjusted to CxHxW, as shown in Figure 3 . Among them, the WFE structure is as shown in Figure 5As shown, three different pooling operations are adopted to obtain three weight feature maps, and the weight feature maps are superimposed to calculate the features to obtain the global representative feature map W υ The three pooling operations are global depth average pooling (GDAP) and global depth maximum pooling (GDMP), and the channel self-attention pooling (context modeling result CM C2S In this section, the abstract definition is Y CM The global representative feature map W υ is expressed as:

[0066] W υ = Y Conv δ(Y CM X υ , Y GDAP X υ , Y GDMP X υ )

[0067] Where Y CM , Y GDAP , Y GDMP represent three pooling operations respectively; X υ represents the input feature layer, υ∈{L * , M * , S *}; δ(·) represents the operation of connecting feature maps; Y Conv represents CBS-R; W υ represents the final global representative weight feature obtained by input.

[0068] As Figure 3 , the function of FRA is to perform feature calculation and re-weighted fusion on the obtained global representative feature map and . Among them, is redefined as λ, μ, τ, expressed as:

[0069]

[0070] λ η , μ η , τ η represent the weights with position index η in the corresponding global representative feature map.

[0071] After superimposing the three weight feature value maps, linear transformation (such as 1x1 convolution) is performed. Then, in the channel dimension, we use the Softmax function to recalculate, and the result is λ r , μ r,τ r Relationship expression:

[0072]

[0073] After obtaining the recalculated weights, the final feature aggregation reweighting is performed. The fusion method is: multiplying the adjusted and the corresponding recalculated weight feature map r ,μ r ,τ r , and then adding the three reweighted feature layers to obtain the final AL, the aggregation expression is:

[0074]

[0075] wherein, indicates that each feature map with the channel dimension as the axis needs to be multiplied by the recalculated weight feature map; (1→C) indicates that the number of feature maps is C; indicates the obtained aggregated reweighted feature, and ω indicates the output layer of any scale in (X L , X M , X S ) as DL.

[0076] In order to verify the effectiveness of the remote sensing rotation detection network based on context feature adaptive weighting fusion described in the application, the following will be compared in detail through experiments.

[0077] The experimental environment is: the operating system is Ubuntu 20.04, the deep learning framework is Pytorch 1.10, and the Python version is 3.9. The application compares the detection results of the model on the remote sensing image dataset DOTA-v1.0 (taking the mainstream mAP50 and FPS as the evaluation index, the higher the result, the higher the precision and speed, and the better the detection performance of the model). The mainstream deep learning-based rotation target detection method in recent years is selected for comparison, specifically:

[0078] BBAVectors: the method proposed by J et al., reference document “Yi J, Wu P, Liu B, et al. Oriented object detection in aerial images with box boundary-aware vectors [C] / / Proceedings of the IEEE / CVF Winter Conference on Applications of Computer Vision. 2021: 2150-2159.”

[0079] AOPG: Method proposed by G et al., reference "Cheng G, Wang J, Li K, et al. Anchor-free oriented proposal generator for object detection [J]. IEEE Transactions on Geoscience and Remote Sensing, 2022, 60: 1-11."

[0080] G-Rep: Method proposed by L et al., reference "Hou L, Lu K, Yang X, et al. G-rep: Gaussian representation for arbitrary-oriented object detection [J]. Remote Sensing, 2023, 15(3): 757."

[0081] S2A-Net: Method proposed by J et al., reference "Han J, Ding J, Li J, et al. Align deep features for oriented object detection [J]. IEEE Transactions on Geoscience and Remote Sensing, 2021, 60: 1-11."

[0082] ReDet: Method proposed by J et al., reference "Han J, Ding J, Xue N, et al. ReDet: A Rotation-Equivariant Detector for Aerial Object Detection. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2786-2795"

[0083] Oriented R-CNN: Method proposed by X et al., reference "Xie X, Cheng G, Wang J, et al. Oriented R-CNN for object detection [C] / / Proceedings of the IEEE / CVF International Conference on Computer Vision. 2021: 3520-3529."

[0084] GWD: Method proposed by X et al., reference "Yang X, Yan J, Ming Q, et al. Rethinking rotated object detection with gaussian wasserstein distance loss [C] / / International conference on machine learning. PMLR, 2021: 11830-11841."

[0085] FDOL: Method proposed by S et al., reference "Zheng S, Wu Z, Xu Y, et al. Learning Orientation Information From Frequency-Domain for Oriented Object Detection in Remote Sensing Images [J]. IEEE Transactions on Geoscience and Remote Sensing, 2022, 60: 1-12."

[0086] KFIoU: Method proposed by X et al., reference "Yang X, Zhou Y, Zhang G, et al. The KFIoU loss for rotated object detection [J]. arXiv preprint arXiv: 2201.12558, 2022."

[0087] The test results are shown in Table 1. Under single-scale training and single-scale testing, the DAG-YOLO proposed by the present application not only has the best detection performance in the list, achieving 77.75%, but also has an FPS of 25.84, far exceeding other rotated object detection algorithms in the list, achieving dual advantages of precision and speed. In addition, the present application achieves the best detection performance on PL, BD, SV, LV, SH, ST, HA and SP in the 15 categories of the DOTA-v1.0 test set, and also Figure 6 shows the detection effect diagram of the present application on the DOTA-v1.0 test set, further indicating that the method of the present application can accurately identify and locate the target in RSI detection, especially for many small targets arranged densely.

[0088] Table 1 Comparison of the latest methods based on the DOTA-v1.0 test data set

[0089]

[0090] It should be noted that the above examples are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that the technical solutions recorded in the foregoing examples can be modified, or some or all of the technical features can be replaced by equivalents; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A remote sensing rotation detection method based on contextual feature adaptive weighted fusion, characterized in that... Includes the following steps: (1) The angles in the OBB tags are encoded into 180 results using the one-dimensional Gaussian angle label encoder GALE and all of them are retained as the category labels of the angles, which contain 180 categories. At the same time, the angle classification loss ACL is designed. (2) Combine the one-dimensional Gaussian angle label encoder in (1) with the YOLOv5 detection network to construct the rotation detector baseline GLR-YOLO, and incorporate the angle classification loss ACL into the total loss of YOLOv5 to constrain and optimize the learning of GLR-YOLO parameters; (3) To address the problem that the complex target distribution in remote sensing image RSI makes feature extraction difficult, a dual-branch context adaptive modeling module DBCAM was designed. The feature space and channel dependencies are obtained through the self-attention pooling context modeling method SPCM, and the dual-path weighted fusion method DPWF is designed to fully capture the dual dependencies between channels and space. DBCAM includes two branching structures, one above the other. The modeling formula for the upper branch S2C includes: Where X represents the abstract feature layer. Γ represents the context modeling result of a feature map, while Γ represents the result of context modeling for each feature map in X. Similarly, the modeling formula for the lower branch C2S includes: Where Y represents the redefined abstract feature layer. Y represents the context modeling result of a feature map, while Ψ represents the result of context modeling for each feature map in Y. By modeling the global context, channel and spatial dependencies are fully captured, and the two-branch structure of feature fusion weighting is combined to provide useful semantic information to supplement the entire input feature. (4) To address the issue of inconsistent features among the three output layers of the detection network, an adaptive global feature aggregation and reweighting module (AGFAR) was designed. For each output layer, the size of the other two output layers was adjusted to the size of that layer using the designed feature shape adjustment module (FSA). Then, the proposed feature aggregation and reweighting module (FARM) was used to extract and reweight the feature maps of the three feature layers. The calculated weights were then adaptively multiplied by each output feature layer to complete feature reweighting. Finally, the reweighted feature layers were added together to complete feature aggregation. The FSA module adjusts the size of the three output layers to ensure their feature map positions are consistent, thereby obtaining the most useful information from the same location. After adjustment, the output layer (X) is redefined. L ,X M ,X S ) was adjusted to (X L* ,X M* ,X S* The FARM module includes Weighted Feature Extraction (WFE) and Feature Aggregation and Reweighting (FAR). After FSA adjustment, three feature layers are obtained. The same WFE is performed on each adjusted feature layer to obtain their respective weighted feature maps. The three weighted feature maps are superimposed and then linearly transformed. Then, the Softmax function is used to recalculate the weights in the channel dimension. After obtaining the recalculated weights, FAR is used for the final feature aggregation and reweighting.

2. The method according to claim 1, characterized in that, Step (1) constructs an angle label encoder GALE using a one-dimensional Gaussian function and a five-parameter representation. This encoder retains 180 result values ​​conforming to a one-dimensional Gaussian distribution for each target's angle label, used for subsequent loss function calculations. Simultaneously, the angle classification loss ACL is proposed using the BCEWithLogitsLoss function, with the expression L... θ for: Where n represents the matched positive samples, C represents that each sample has C classes, and y c x represents the ground truth label (GT) of class c in sample n, which is 0 or 1. c θ represents the probability that sample n, predicted by the network model, belongs to class c. P θ represents the 180 channel values ​​predicted by the network model for each sample. GL This represents the result of performing one-dimensional Gaussian encoding on the angles in the ground plane.

3. The method according to claim 1, characterized in that, In step (2), GLR-YOLO is an RSI neighborhood rotating target detector constructed by combining GALE and ACL with the classic one-stage horizontal target detector YOLOv5.

4. The method according to claim 1, characterized in that, In step (3), to address the problem of complex target distribution in remote sensing image RSI leading to difficulty in feature extraction, a dual-branch context adaptive modeling module (DBCAM) was designed. Two modeling methods and feature transformation fusion paths were constructed using self-attention pooling. DBCAM includes two upper and lower branch structures. The modeling formula for the upper branch S2C includes: Where X represents the abstract feature layer. Γ represents the context modeling result of a feature map, while Γ represents the result of context modeling for each feature map in X. Similarly, the modeling formula for the lower branch C2S includes: Where Y represents the redefined abstract feature layer. Y represents the context modeling result of a feature map, while Ψ represents the result of context modeling for each feature map in Y. Both the S2C and C2S branches employ the same dual-path weighted fusion method. DPWF fully captures the dual dependencies between channels and space, namely: WITH S2C =XSigmoid(W v12 (ReLU(LN(W v11 (Γ)))))+W v22 (ReLU(LN(W v21 (Γ)))) Z C2S =YOU S1 =Sigmoid(U s12 (ReLU(U s11 (Ψ))))+U s22 (ReLU(U s21 (Ψ))) Among them, W v11 W v12 W 21 W 22 All are convolutions with a kernel size of 1×1, U s11 U s21 U represents a convolution with a kernel size of 7×7. s12 U s22 This represents a convolution with a kernel size of 3×3; The results of the two branches are added together to obtain the final output of DBCAM: Z=Z S2C +Z C2S By modeling the global context, channel and spatial dependencies are fully captured, and the two-branch structure of feature fusion weighting provides useful semantic information to supplement the entire input features.

5. The method according to claim 1, characterized in that, In step (4), to address the issue of inconsistent features among the three output layers of the detection network, an adaptive global feature aggregation and reweighting module (AGFAR) is designed, which includes a feature shape alignment module (FSA) and a feature aggregation and reweighting module (FARM). When each output layer is used as the aggregation-dominant layer (DL), the other two output layers are used as aggregation-supplementary layers (SLs) to participate in feature aggregation and reweighting. For FSA, its function is to adjust the size of the three output layers to ensure that their feature mapping positions are consistent, thereby obtaining the most useful information from the same location. After adjustment, the output layer (X) is redefined. L ,X M ,X S ) was adjusted to The FARM module includes Weighted Feature Extraction (WFE) and Feature Aggregation and Reweighting (FAR). After FSA adjustment, three feature layers are obtained. The same WFE is performed on each adjusted feature layer to obtain its respective weighted feature map, represented as follows: W υ =Y Conv d(Y CM X υ ,Y GDAP X υ ,Y GDMP X υ ) Among them Υ CM ,Υ GDAP ,Υ GDMP These represent three pooling operations; X υ Denotes the input feature layer, υ∈{L} * M * ,S * };δ(·) represents the operation of connecting feature maps;Υ Conv Indicates CBS-R; W υ This represents the final globally representative weight feature obtained from the input; Then Redefined as λ,μ,τ, expressed as: After superimposing the weight maps of the three weighted features, a linear transformation is performed; then, the Softmax function is used to recalculate the result in the channel dimension, yielding λ. r ,μ r ,τ r , Relational expressions: After obtaining the recalculated weights, the final feature aggregation and reweighting are performed using FAR; the aggregation expression is: in, This indicates that each feature map along the channel dimension needs to be multiplied by a recalculated weighted feature map; (1→C) indicates that the number of feature maps is C; The obtained aggregated weighted features are represented by ω, which is denoted as (X). L ,X M ,X S The output layer of any scale in the ) is used as the DL.

Images