Remote sensing rotation detection network based on context feature adaptive weighted fusion

Abstract

The invention discloses a remote sensing rotation detection network based on context feature adaptive weighted fusion. The method mainly comprises the following steps: (1) designing a one-dimensional Gaussian angle label encoder (GALE) to encode angles in an OBB label, and proposing angle classification loss (ACL); (2) GALE, ACL and YOLOv5 are combined, and a rotating target detector baseline GLR-YOLO is constructed; (3) designing a double-branch context adaptive modeling module (DBCAM) to perform context feature modeling, conversion and fusion so as to reconstruct a feature extraction network; and (4) providing an adaptive global feature aggregation and reweighting module (AGFAR), constructing an adaptive detection head (ADH) in combination with the detection head, and solving the problem of output interlayer feature inconsistency. The method 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, national defense safety and the like.

Description

Technical Field

[0001] The invention relates to a remote sensing rotation 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 Art

[0002] Object detection is a challenging task in computer vision, which aims to identify and classify objects in images. In the past decade, object detection methods based on deep learning and convolutional neural networks have been continuously innovated. Among them, the Rotating Object Detection (ROD) task, a more professional task in remote sensing image (RSI) detection, has attracted more and more attention from researchers.

[0003] RSI refers to the non-contact acquisition of object information at a great distance through different imaging sensors, and the obtained visible light images have the characteristics of high resolution, clear targets, and optional observation range. Remote sensing ROD provides a new detection method for target detection, and has important research value in the fields of building warning, traffic monitoring, military defense security, etc. Different from the traditional target detection technology that mainly uses horizontal bounding boxes (HBB), in RSI, due to its wide coverage, many target categories in the image and uneven size distribution of different categories, and the image background carries a lot of noise, the imaging is complex and the target is easily isolated, etc., it is difficult for the target detection algorithm in natural scenes to achieve good results in remote sensing images. In addition, the overlap between the two HBBs in RSI may be very large, making the non-maximum suppression (NMS) based on HBB unsuitable for processing oriented objects. Therefore, the study of ROD method using oriented bounding boxes (OBB) is necessary for accurate target detection in RSI. In view of the above problems, the present invention takes the target detection task in the RSI field as the research object, and proposes a remote sensing rotation detection network based on adaptive weighted fusion of contextual features, aiming to improve the detection accuracy and detection effect of targets in RSI, especially small targets, while ensuring the detection speed. Summary of the invention

[0004] The present invention proposes a remote sensing rotation detection network (DAG-YOLO) based on adaptive weighted fusion of contextual features, which mainly consists of three parts: a one-dimensional Gaussian angle label encoder (1D Gaussian Angle Label-Encoder, GALE), a dual branch context adaptive modeling module (Dual Branch Context Adaptive Modeling module, DBCAM) and an adaptive global feature aggregation and reweighting module (Adaptive Global Feature Aggregation and Reweighting module, AGFAR). Specifically, GALE is mainly responsible for using a 1D Gaussian function to encode 180 result values ​​from the angle label in the OBB label, where the result indexed as the angle label value is 1, and the remaining values ​​conform to the 1D Gaussian distribution; DBCAM uses channel self-attention and spatial self-attention to perform context-adaptive modeling on the network input feature layer, fully capturing the dual dependencies of the feature layer channel and space, improving the utilization of effective features, and thereby reconstructing the feature extraction network; AGFAR combines the network detection head to construct an adaptive detection head, which aggregates all output feature layers and re-weights and fuses them, while solving the problem of inconsistent features between different output layers, and achieving accurate prediction of large, medium, and small targets. The coordinated use of the above three parts enables the present invention to construct an efficient and accurate rotating target detection network, which can achieve accurate positioning and identification of targets in RSI.

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

[0006] (1) Use the one-dimensional Gaussian angle label encoder (GALE) to encode the angles in the OBB label into 180 results and retain all of them as the category labels of the angles (including 180 categories), and design the angle classification loss (ACL);

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

[0008] (3) Aiming at the problem that the complex distribution of targets in remote sensing images (RSI) makes it difficult to extract features, a dual-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 dependency, and the dual dependency between channel and space is fully captured by the designed double path weighted fusion method (DPWF).

[0009] (4) To address the problem of inconsistent features among the three output layers of the detection network, an adaptive global feature aggregation and reweighting module (AGFAR) is designed. For each layer of output, the output sizes of the other two layers are adjusted to the size of this layer through the designed Feature Shape Alignment module (FSA). Then, the proposed Feature Aggregation Reweighting module (FAR) is used to extract and reweight the weight feature maps of the three feature layers. The calculated weights are adaptively multiplied by each output feature layer to complete feature reweighting. Finally, the reweighted feature layers are added together to complete feature aggregation.

[0010] Compared with the prior art, the present invention has the following beneficial effects:

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

[0012] 2. The constraints of the ACL designed by the present invention on the angle classification prediction not only ensure the accuracy of the angle prediction, but also balance the total loss of the rotation detector, thereby enhancing the robustness and stability of the model;

[0013] 3. The present invention adopts DBCAM to reconstruct the feature extraction network, performs context modeling and feature capture fusion in a self-attention pooling manner, and reconstructs the feature extraction network while controlling the complexity of the model, thereby enhancing the network's context feature extraction capability and the ability to locate targets under high thresholds;

[0014] 4. The present invention designs an adaptive detection head in combination with AGFAR, which re-weights 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 present invention adopts the rotation detector DAG-YOLO constructed by a one-stage network, which not only maintains excellent performance in speed, but also obtains a stable improvement in detection performance, proving its advantage and scalability in rotation detection tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is the overall model block diagram 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 re-weighting module of the present invention;

[0019] Figure 4 This is a structural diagram of a characteristic 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 a remote sensing image. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings. It should be understood that the specific implementation methods described herein are only used to explain the present invention and are not used to limit the present invention.

[0023] like Figure 1 As shown, a rotation target detection method for RSI is mainly divided into three structures: Backbone, Neck and Head. Among them, the proposed DBCAM is fused with the C3 structure to form a C3D structure, and then the Backbone and Neck are reconstructed for feature extraction; the proposed AGFAR is combined with the Head to construct an adaptive detection head (Adaptive Detect Head, ADH), which solves the problem of inconsistent features between output layers; the network's final predicted 180 angle category scores and GALE encoding results use ACL to calculate the loss, and combine other losses to optimize the network parameters. The following is a further description of each part of the present invention:

[0024] 1. Rotation Detector Baseline GLR-YOLO

[0025] GLR-YOLO is a rotation detector built with YOLOv5 as the baseline, including the core improvement innovations, namely GALE and ACL. The idea of ​​GALE is to transform the angle regression task into an angle classification task, and use the 180 channel values ​​predicted by the network to fit the result value after the angle label encoding in the OBB. The specific process is:

[0026] First, the original target labels (x1, y1, x2, y2, x3, y3, x4, y4) marked with arbitrary quadrilaterals in RSI are converted into OBB labels (x, y, w, h, θ). The OpenCV representation (D oc ) to Long-Edge representation (D le ), then apply GALE to D le The angle θ le Encode, and 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 proposed and introduced ACL. ACL depends on the BCEWithLogitsLoss function, which is expressed as:

[0030]

[0031] Among them, 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 that sample n belongs to category c predicted by the network model.

[0032] Use ACL to obtain the angle classification loss L θ The formula is:

[0033]

[0034] Among them, θ P Represents the 180 channel values ​​predicted by the network model for each sample; θ GL It represents the result of one-dimensional Gaussian encoding of the angles in GT using Algorithm 1.

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

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

[0037] Since the two branch structures are executed simultaneously from input to output, we abstract the proposed method into two steps: first, adaptive feature context modeling; second, feature transformation and fusion. All dimensions of the feature layer change as follows Figure 6 shown.

[0038] The adaptive feature context modeling process of S2C is as follows: the feature layer obtained after the image is extracted through multiple convolutional layers is abstracted into C feature maps of size H×W, 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 The position index of the feature information in the g feature map is i. The final output is defined as Z S2C Therefore, the modeling formula for spatial self-attention pooling to obtain contextual features is defined as:

[0041]

[0042] Where j,m represents 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 map and shared weights, the global context feature modeling formula for the entire feature layer can be obtained:

[0043]

[0044] in, Represents the modeling result of the contextual features of a feature map.

[0045] Adaptive feature context modeling process of C2S: Different from the S2C modeling process, we consider the feature vector from the channel dimension, redefine the input feature vector as Y, and abstract it into P feature maps of size C×1, so that the number of positions on a feature map is C. Therefore, we can define the feature map as Where t represents the position index. Further, Y is expressed as:

[0046]

[0047] where y ht Represents the feature information on the h feature map with position index t, and defines the output as Z C2S Finally, we use channel self-attention pooling to model the contextual features, which is expressed as:

[0048]

[0049] Among them, f,n represents the enumeration of all possible positions in a feature map; GAP and GMP represent global average pooling and global maximum pooling respectively. f represents the global shared channel attention weight; W k2 Y represents the initialization 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] Among them, h ranges from 1 to P, indicating that contextual feature modeling is performed on each feature map of the abstract feature layer Y.

[0052] Feature transformation and fusion: For both the S2C branch and the C2S branch, we use a dual-path feature transformation module to correspond to the two fusion weighted modes. Through feature transformation and dual-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 difficulty of optimization and improve generalization ability, we use layer normalization (LN) before the ReLU function to fully capture the feature dependencies between channels. In order to ensure the consistency of 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 =XSigmoid(W v12 (ReLU(LN(W v11(Γ)))))+W v22 (ReLU(LN(W v21 (Γ))))

[0055] Among them, W v11 ,W v12 ,W 21 ,W 22 All of them are convolutions with a kernel size of 1×1.

[0056] In the proposed C2S branch, we first use a large kernel convolution with a kernel size of 7×7, and then use the ReLU function and a convolution with a kernel size of 3×3 to fully capture the spatial dependency of features. In order to maintain the consistency of feature distribution, the same dual-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] Among them, U s11 ,U s21 represents a convolution with a kernel size of 7×7, U s12 ,U s22 Represents a convolution with a kernel size of 3×3.

[0059] Add the fusion results of the two branches to get the final output of DBCAM:

[0060] Z=Z S2C +Z C2S

[0061] Through global context modeling, channel dependency and spatial dependency are fully captured, and combined with the two branch structures (S2C and C2S) of feature fusion weighting, useful semantic information supplement is provided for the entire input feature.

[0062] 3. Adaptive Global Feature Aggregation and Reweighting 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 objectives between different scale features. Figure 3As shown in the figure, in order to solve the problem of inconsistent features between FPN output layers and fully learn the feature information between feature layers of different scales, AGFAR is proposed, which mainly consists of two parts: Feature Shape Alignment module (FSA) and Feature Aggregation Reweighting Module (FARM). In order to obtain the final aggregation feature layer (AL) of any scale, we define the input feature layer of this scale (such as P5) as the dominant layer of polymerization features (DL), and define the other two input feature layers (P4 and P3) as supplementary layers of polymerization features (SLs).

[0064] like Figure 4 As shown in Figure 2, 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, since the scale of the output layer as DL is different, although the method of feature aggregation and reweighting is the same, the method of adjusting the feature shapes of the two SLs to the DL shape is different, such as Figure 4 (a)(b)(c) show the alignment scheme of feature shape and DL respectively. This scheme uses convolution layer (Conv) to adjust the number of channels, and the maximum pooling layer (MP) and linear interpolation upsampling layer (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 inconsistent features between output layers. It has two important components: Weight Feature Extraction (WFE) and Feature Aggregation Reweighting (FAR). The size of the feature layer used as DL in the three output layers of the image is defined as C×H×W. After the FSA module, all three output layers are adjusted to C×H×W, such as Figure 3 As shown. Among them, the WFE structure is as follows Figure 5As shown in Figure 1, three different pooling operations are used to obtain three weighted feature maps, which are then superimposed to perform feature calculations to obtain a global representative feature map W. υ , the three pooling operations are: Global Depth Average Pooling (GDAP) and Global Depth Maximum Pooling (GDMP), as well as the channel self-attention pooling (context modeling result CM C2S , which is abstractly defined in this section as CM ), the global representative feature map W υ It is expressed as:

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

[0067] Among them CM ,Υ GDAP ,Υ GDMP Represent three pooling operations respectively; X υ represents the input feature layer, υ∈{L * ,M * ,S *}; δ(·) represents the operation of connecting feature graphs; Υ Conv Indicates CBS-R; W υ Represents the final global representative weight feature obtained from the input.

[0068] like Figure 3 The function of FRA is to obtain the global representative feature map and Perform feature calculation and re-weighted fusion. is redefined as λ, μ, τ and expressed as:

[0069]

[0070] λ η ,μ η ,τ η represents the weight at position index η in the corresponding global representative feature map.

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

[0072]

[0073] After obtaining the recalculated weights, the final feature aggregation reweighting is performed. The fusion method is: and the corresponding recalculated weight feature map λ r ,μ r ,τ r Multiply them together, and then add the three re-weighted feature layers to get the final AL. The aggregation expression is:

[0074]

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

[0076] In order to verify the effectiveness of the remote sensing rotation detection network based on context feature adaptive weighted fusion described in the present invention, a detailed comparison will be carried out through experiments below.

[0077] The experimental environment is: operating system Ubuntu20.04, deep learning framework Pytorch 1.10, Python version 3.9. The present invention compares the detection results of the model on the remote sensing image dataset DOTA-v1.0 (using mainstream mAP50 and FPS as evaluation indicators, the higher the result, the higher the accuracy and speed, and the better the detection performance of the model). The mainstream rotation target detection methods based on deep learning in recent years are selected for comparison, specifically:

[0078] BBAVectors: The method proposed by Yi et al., reference "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: The 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: The 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: The 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: A method proposed by J et al., reference "Han J, Ding J, Xue N, et al. ReDet: ARotation-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: The 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: The 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: The 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: The 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 proposed DAG-YOLO not only achieves the best detection performance in the list, achieving 77.75%; at the same time, the FPS reaches 25.84, far exceeding other rotation target detection algorithms in the list, achieving the dual advantages of accuracy and speed. In addition, it achieves the best detection performance in PL, BD, SV, LV, SH, ST, HA and SP among the 15 categories of the DOTA-v1.0 test set. Figure 6 The detection effect diagram of the present invention on the DOTA-v1.0 test set is also displayed, which further shows that the method of the present invention can accurately identify and locate targets in RSI detection, especially many densely arranged small targets.

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

[0089]

[0090] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. However, 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 invention.

Claims

1. A multi-task image restoration method based on multi-scale degraded representation, characterized in that The following steps are involved: (1) Use the one-dimensional Gaussian angle label encoder (GALE) to encode the angles in the OBB label into 180 results and retain all of them as the category labels of the angles (including 180 categories), and design the angle classification loss (ACL); (2) The one-dimensional Gaussian angle label encoder in (1) is combined with the YOLOv5 detection network to construct the rotation detector baseline GLR-YOLO, and the angle classification loss (ACL) is incorporated into the YOLOv5 total loss to constrain and optimize the learning of GLR-YOLO parameters; (3) Aiming at the problem that the complex distribution of targets in remote sensing images (RSI) makes it difficult to extract features, a dual-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 dependency, and the dual dependency between channel and space is fully captured by the designed double path weighted fusion method (DPWF). (4) To address the problem of inconsistent features among the three output layers of the detection network, an adaptive global feature aggregation and reweighting module (AGFAR) is designed. For each layer of output, the output sizes of the other two layers are adjusted to the size of this layer through the designed Feature Shape Alignment module (FSA). Then, the proposed Feature Aggregation Reweighting module (FAR) is used to extract and reweight the weight feature maps of the three feature layers. The calculated weights are adaptively multiplied by each output feature layer to complete feature reweighting. Finally, the reweighted feature layers are added together to complete feature aggregation.

2. The method according to claim 1, characterized in that: The step (1) uses a one-dimensional Gaussian function (1DGaussian) and a five-parameter representation (D le ) constructed an angle label encoder (GALE), which retains 180 result values ​​that conform to the one-dimensional Gaussian distribution for the angle label of each target by encoding, which is used for the calculation of the subsequent loss function. At the same time, the angle classification loss (ACL) is proposed using the BCEWithLogitsLoss function, and its expression L θ for: Among them, n represents the matched positive sample, C represents each sample has C categories, and y c represents the GT label (0 or 1) of category c in sample n, x c represents the probability that the sample n predicted by the network model belongs to category c, θ P Represents the 180 channel values ​​predicted by the network model for each sample, θ GL It represents the result of one-dimensional Gaussian encoding of the angles in GT using Algorithm 1.

3. The method according to claim 1, characterized in that The GLR-YOLO in step (2) is a high-precision RSI field rotation target detector constructed by combining GALE and ACL with the classic one-stage horizontal target detector YOLOv5; it not only solves the detection difficulty caused by the complex target distribution in RSI, but also transforms the angle regression prediction into classification prediction, solves the boundary problem commonly existing in the rotation detector, greatly reduces the optimization difficulty of the rotation detector, and improves the detection accuracy in various difficult scenarios.

4. The method according to claim 1, characterized in that: In the step (3), a dual-branch context adaptive modeling module (DBCAM) is designed to solve the problem that the complex distribution of targets in the remote sensing image (RSI) makes it difficult to extract features. Two modeling methods and feature conversion fusion paths are constructed by self-attention pooling. The DBCAM includes two upper and lower branch structures. The modeling formula of the upper branch S2C includes: Among them, X represents the abstract feature layer, represents the context modeling result of a feature map, and Γ represents the result of context modeling for each feature map in X; Similarly, the modeling formula of the lower branch C2S includes: Among them, Y represents the redefined abstract feature layer, represents the context modeling result of a feature map, and Ψ represents the result of context modeling for each feature map in Y; Both S2C and C2S branches use the same double path weighted fusion (DPWF) to fully capture the dual dependency between channels and spaces, which are: 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 of them are convolutions with a kernel size of 1×1, U s11 ,U s21 represents a convolution with a kernel size of 7×7, U s12 ,U s22 Indicates a convolution with a kernel size of 3×3; Add the fusion results of the two branches to get the final output of DBCAM: Z=Z S2C +Z C2S Through global context modeling, channel dependency and spatial dependency are fully captured, and combined with the two branch structures (S2C and C2S) of feature fusion weighting, useful semantic information supplement is provided for the entire input feature.

5. The method according to claim 1, characterized in that In the step (4), an adaptive global feature aggregation and reweighting module (AGFAR) is designed to solve the problem of inconsistent features among the three output layers of the detection network. The core modules of the AGFAR include a feature shape alignment module (FSA) and a feature aggregation reweighting module (FARM). The core idea is that when each output layer is used as an aggregation leading layer (DL), the other two output layers are used as aggregation supplementary layers (SLs) to participate in feature aggregation and reweighting. For FSA, its role 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. After adjustment, the redefined output layer (X L ,X M ,X S ) is adjusted to The FARM module is the core step to solve the problem of inconsistent features between output layers. It has two important components: Weight Feature Extraction (WFE) and Feature Aggregation Reweighting (FAR); the size of the three feature layers after FSA adjustment is, and then, the same WFE is performed on each adjusted feature layer to obtain their respective weight feature maps, which are expressed as: W υ =Y Conv d(Y CM X υ ,Y GDAP X υ ,Y GDMP X υ ) Among them CM ,Υ GDAP ,Υ GDMP Represent three pooling operations respectively; X υ represents the input feature layer, υ∈{L * ,M * ,S * }; δ(·) represents the operation of connecting feature graphs; Υ Conv Indicates CBS-R; W υ Represents the final global representative weight feature obtained by the input; Then Redefine them as λ, μ, τ, expressed as: After superimposing the three weight feature weight maps, we perform a linear transformation (such as 1×1 convolution); then, in the channel dimension, we recalculate using the Softmax function, and the result is Relational expression: After obtaining the recalculated weights, FAR is used to perform the final feature aggregation reweighting to obtain the final AL; the aggregation expression is: in, Indicates that each feature map with channel dimension as axis needs to be multiplied by the recalculated weight feature map; (1→C) indicates that the number of feature maps is C; represents the obtained aggregated weighted features, and ω is represented by (X L ,X M ,X S ) as the output layer of any scale in .

Images