Deep learning-based depression angle sensitive small target remote identification method

Abstract

The invention discloses a depression angle sensitive small target remote identification method based on deep learning, and relates to the technical field of potential sensitive small target identification in a complex environment. For example, long-distance identification and positioning of small targets in a disaster area can be realized in an emergency disaster relief scene. According to the invention, based on a YOLOv8 target detection model, an improved Swin Transform module is introduced, so that the model keeps enough global context information in a feature extraction process, and more differentiated features are extracted through a self-attention mechanism; an RFA module and a lightweight attention mechanism ECA are introduced, a neck FPN structure of an original algorithm is upgraded to RE-FPN, and the attention capability of the network to important features is improved; a small target detection layer is added to enhance the perception ability of the model to spatial feature information; and finally, by using a target detection head Dynamic Head with multiple attention mechanisms, the detection performance of the detection head on a low-resolution target under a complex background is improved, and identification and detection of sensitive small targets such as people, vehicles and important facilities in a complex scene picture are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of sensitive target recognition in complex environments, and specifically to a method for long-distance intelligent recognition of sensitive small targets at a depression angle based on deep learning. Background Art

[0002] With the rapid development of deep learning theory and models, target detection technology based on deep learning has become more mature, and target detection has gradually become a research hotspot. It has been widely used in medicine, industry, transportation, emergency management, emergency relief and military confrontation. In important fields such as military and emergency rescue, target detection occupies an extremely important position. Modern warfare and modern emergency management are increasingly dependent on cutting-edge technology. Taking war as an example, information intelligence confrontation has also become the main type of war today. Target detection technology has broad application prospects and extremely important roles in the military and emergency fields. It has become an important technical support means for realizing the intelligence and informatization of war and emergency management, and is also an important part of military and emergency rescue operations. Accurate and efficient small target detection in complex battlefield environments or emergency rescue environments can help rescuers or troops quickly detect important sensitive small targets, and obtain the location information and category information of the target in a timely manner, which is conducive to seizing the initiative and quickly and accurately implementing subsequent tasks. Moreover, target detection is the basis of target tracking, and its detection effect plays a decisive role in subsequent target tracking and battlefield situation analysis. Among many reconnaissance or emergency rescue methods, drones, with their unmanned, low-cost, and flexible features, have replaced large manned reconnaissance aircraft and become the most important reconnaissance and rescue force at the front of the battlefield. Drones cruise at ultra-low or high altitudes, shoot battlefield conditions below from a bird's-eye view, collect real-time scene images in complex environments, and detect sensitive targets. Due to the long shooting distance, there are a large number of small-sized targets in these images, and the small targets captured from a bird's-eye view are different from those captured from an ordinary head-on view. Affected by the shooting angle, the targets captured at a bird's-eye view will undergo posture shifts and other changes. In addition, the image data background is highly complex, the target appearance is blurred, and the distance increases the difficulty of detection. In addition, the image acquisition camera carried is generally movable and rotatable, and image data is collected from multiple angles and scales. The use scenarios are more extensive, which requires the detection algorithm to have strong generalization capabilities and be able to accurately locate and detect targets when the size and posture of distant targets change. In this context, how to accurately detect sensitive potential small targets in image data captured from a bird's-eye view in complex scenes is an important issue that needs to be solved urgently.

[0003] In the past, target detection tasks mainly relied on manual inspection methods. For key tasks, it was necessary to configure a group of people to monitor the screen. Especially when monitoring long-distance video screens to detect small targets, manual visual methods will inevitably have problems such as omissions and misjudgments. Some traditional computer vision detection algorithms rely on manually designed feature extraction methods, which cannot capture higher-dimensional semantic information, and the detection effect is often unsatisfactory. With the emergence of deep learning target detection algorithms, target detection technology has greatly improved the detection effect of common general targets, but there is still a lack of detection models for sensitive small targets. On the one hand, due to the confidentiality and security of various data in complex environments, there are almost no public related complex environment target data sets at home and abroad, and sensitive target data sets are relatively scarce; on the other hand, the real environment is complex and changeable, and the detection and recognition of sensitive small targets are easily affected by the environment, resulting in a lack of accuracy. Therefore, most detection models cannot meet the requirements of high-precision and low-miss detection sensitive target detection tasks.

[0004] Chinese invention patent application number CN202110701296.X discloses a method for detecting and classifying sea surface military targets based on an improved EfficientDet network, which uses a model based on a baseline network to improve a backbone network and a feature fusion network to detect sea surface military targets. This patent can effectively identify targets such as sea surface ships, but cannot effectively detect other types of targets such as people and vehicles in complex environments, and the detection model cannot be deployed on unmanned equipment to perform real-time detection of the battlefield environment.

[0005] In summary, in order to address the problems of limited recognition methods for sensitive small targets in scene pictures taken from a bird's-eye view in complex emergency environments, low long-distance recognition accuracy, poor real-time performance, and a single detectable target category, it is particularly necessary to develop a long-distance recognition method for sensitive small targets at a bird's-eye view based on deep learning. Summary of the invention

[0006] In view of the shortcomings of the prior art, the purpose of the present invention is to provide a method for long-distance intelligent recognition of small targets sensitive to depression angles based on deep learning, so as to realize rapid and accurate detection of sensitive targets in scene pictures taken from an air-to-ground bird's-eye view in complex emergency environments, and obtain the category information and location information of the target so that the commander of the own side can make reasonable decisions quickly, thereby seizing the initiative and defeating the enemy.

[0007] In order to achieve the above-mentioned purpose, the present invention is implemented through the following technical scheme: a sensitive target recognition method under depression angle based on deep learning, wherein the basic network YOLOv8 target detection model consists of three parts: a backbone network, a neck network and a head network. Firstly, an improved Swin Transformer structure is introduced into the original backbone network to enhance the feature extraction capability of the backbone network for input images; then, the original FPN structure is replaced with a lighter and more efficient residual feature fusion pyramid structure RE-FPN, and a small target detection layer is added to improve the network's detection capability for objects of different scales, especially small target objects; finally, a dynamic detection head Dynamic Head with a multiple attention mechanism is introduced to make the model pay more attention to dense small target areas and extract more small target features to improve the detection speed and detection accuracy of sensitive targets such as people, vehicles, tanks, warships, radars, etc. in scene images, and obtain the category information and position coordinate information of sensitive targets in real-time collected images.

[0008] Preferably, the backbone network is used to perform feature extraction operations on the input model image, and an improved Swin Transformer structure is introduced into the C2f feature extraction module in the original backbone network to form a C2f-ST feature extraction module, and the C2f module is used to replace all C2f modules in the original backbone network to obtain a new feature extraction backbone network, enhance the feature extraction capability of the network, better retain context information, and improve the detection accuracy of small targets in the input image.

[0009] Preferably, the neck network is responsible for fusing the extracted feature maps, introducing the RFA (Residual Feature Augmentation) residual feature enhancement module, using residual branches, improving the original FPN (Feature Pyramid Networks) structure in YOLOv8, injecting contextual information of different spaces into it to improve the feature expression of high-level features, and enhancing the network's ability to detect small targets. A lightweight attention mechanism ECA module is inserted into the FPN structure to improve the network's ability to focus on important features with little impact on the network parameters and computational complexity.

[0010] Preferably, the head network is composed of four detection heads, that is, a small target detection layer is added to the original neck network, and a detection head for processing large-size feature maps is added to the head network, so that on the basis of the original 20x20, 40x40, and 80x80 features to be detected, a group of 160x160 feature maps to be detected are added, which helps the network to better capture shallow semantic information, and uses multi-scale features to help the model better process targets of different sizes.

[0011] Preferably, the head network complexly processes the detection feature map and outputs sequence vectors of detection results of different scales. We use Dynamic Head to replace the original detection head, introduce a target detection head method with multiple attention mechanisms, further process the output feature map, and introduce three attention mechanisms in different dimensions to improve its adaptability in scale perception, spatial position and multi-task, significantly improving the expression ability of the detection head.

[0012] Preferably, the Swin Transformer module is composed of a layer normalization module (LN), a window-based multi-head self-attention module (W-MSA), a shifted window-based multi-head self-attention module (SW-MSA), and a two-layer multi-layer perceptron (MLP) module including a GELU nonlinear layer. The specific process of inputting feature data from the l-1 layer to the l+1 layer is shown as follows:

[0013]

[0014] LN can make the network better integrated and prevent the network from overfitting. However, using LN layers in CNN will destroy the learned sample features. Therefore, in order to preserve the generalization ability of the model while protecting CNN's ability to learn sample features, the LN in the input part of the cascaded Swin Transformer is removed to make it more suitable for the YOLOv8 network. The specific process of improving the input feature data from the l-1 layer to the l+1 layer is as follows:

[0015]

[0016] The improved Swin Transformer is connected in parallel with the CSP Bottleneck structure in the C2f module to form the C2f-ST feature extraction module. The powerful feature extraction capability of Swin Transformer is utilized to enable the backbone network to better extract the feature information of the input image.

[0017] Preferably, the RFA (Residual Feature Augmentation) residual feature enhancement module reduces the spatial information loss of the feature map extracted by the backbone network by collecting spatial context information, and at the same time improves the performance of the feature pyramid. Considering the aliasing effect that may be caused by interpolation, the ASF (Adaptive Spatial Fusion) adaptive spatial fusion module is introduced to adaptively combine context features, generate a spatial weight map for each feature, and use weights to aggregate context features into M, giving it context information at multiple scales.

[0018] Preferably, the ECA (Efficient Channel Attention) attention mechanism reconstructs the calculation weights of different parts of the feature map during feature fusion, so that the model can improve its attention to the target area. The ECA attention mechanism consists of three steps: compression, excitation and feature recalibration. It greatly reduces the amount of network parameters while avoiding the side effects brought about by the dimensionality reduction operation.

[0019] Preferably, the Dynamic Head dynamic detection head and the attention mechanism unify the detection heads of various targets. By utilizing the attention between feature levels of scale perception, the attention between spatial positions of spatial perception, and the attention between output channels of task perception, the expressiveness of the model target can be greatly enhanced, and the detection of small targets can be more flexible and accurate.

[0020] The scale-aware attention module fuses features of different scales according to their semantic importance as shown in the following equation.

[0021]

[0022] f(·) is a linear function approximated using a 1×1 convolution, σ(x)=max(0,min(1,(x+1) / 2)).

[0023] The spatially aware attention module first uses deformable convolution to learn sparsity and then aggregates cross-layer features at the same spatial location. Where K is the number of sparsely sampled locations, p k +Δp k is the self-learned spatial shift focused on the shifted position of the unique region, Δm k is self-learning at position p k are important scalars at , and are all learned from the input features of the intermediate layer F.

[0024]

[0025] π C (F)·F=max(α 1 (F)·F C +β 1 (F),α 2 (F)·F C +β 2 (F)

[0026] The task-aware attention module dynamically opens or closes feature channels and selects different tasks, where θ(·) = [α 1 ,β 1 ,α 2 ,β 2 ] T, is a hyperfunction that learns to control the activation threshold. It first performs global pooling to reduce the dimension in the L×S dimension, then uses two fully connected layers, a normalization layer, and finally uses a shifted sigmoid function to normalize the output to [-1,1].

[0027] Beneficial effects of the present invention: The present invention realizes customized small target detection of image data taken from an air-to-ground bird's-eye view in a complex emergency environment using an improved YOLOv8 model, obtains real-time category information and location information of sensitive targets in the scene, and provides intelligent support for information decision-making in various fields such as military confrontation and emergency management, with broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments;

[0029] Figure 1 This is a schematic diagram of the improved YOLOv8 target detection network of the present invention;

[0030] Figure 2 The improved Swin Transformer structure diagram of the present invention;

[0031] Figure 3 This is a structural diagram of the improved C2f-ST module of the present invention;

[0032] Figure 4 Schematic diagram of the lightweight residual feature fusion pyramid structure RE-FPN of the present invention;

[0033] Figure 5 Schematic diagram of the structure of the RFA residual feature enhancement module of the present invention;

[0034] Figure 6 This is the structure diagram of the ECA lightweight attention mechanism of the present invention;

[0035] Figure 7 This is a structural diagram of the Dynamic Head of the present invention. DETAILED DESCRIPTION

[0036] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the present invention is further explained below in conjunction with specific implementation methods.

[0037] Reference Figure 1-7The specific implementation of the present invention adopts the following technical scheme: a method for identifying sensitive targets at depression angles based on deep learning, characterized in that its basic network YOLOv8 target detection model consists of three parts: a backbone network, a neck network, and a head network. First, an improved Swin Transformer structure is introduced into the original backbone network to enhance the backbone network's feature extraction capability for input images; then, the original FPN structure is replaced with a lighter and more efficient residual feature fusion pyramid structure RE-FPN, and a small target detection layer is added to improve the network's detection capability for objects of different scales, especially small target objects; finally, a dynamic detection head Dynamic Head with a multiple attention mechanism is introduced to make the model pay more attention to dense small target areas and extract more small target features to improve the detection speed and accuracy of sensitive small targets such as people, vehicles, and important infrastructure in scene images, thereby obtaining sensitive target category information and coordinate information, and providing valuable auxiliary intelligence for military or emergency tasks.

[0038] Specifically, the network functions of each part of the improved YOLOv8 target detection model are as follows:

[0039] (1) The backbone network is used to perform feature extraction operations on the input model image. The improved Swin Transformer structure is introduced into the C2f feature extraction module in the original backbone network to form a C2f-ST feature extraction module, and it is used to replace all C2f modules in the original backbone network to obtain a new feature extraction backbone network, enhance the feature extraction capability of the network, and better retain context information to improve the detection accuracy of small and medium-sized targets in the input image. The Swin Transformer module is composed of a layer normalization module (LN), a window-based multi-head self-attention module (W-MSA), a shifted window-based multi-head self-attention module (SW-MSA) and a two-layer multi-layer perceptron (MLP) module containing a GELU nonlinear layer. The specific process of input feature data from the l-1 layer to the l+1 layer is shown as follows:

[0040]

[0041]

[0042] LN can make the network better integrated and prevent the network from overfitting. However, using LN layers in CNN will destroy the learned sample features. Therefore, in order to preserve the generalization ability of the model while protecting CNN's ability to learn sample features, the LN in the input part of the cascaded Swin Transformer is removed to make it more suitable for the YOLOv8 network. The specific process of improving the input feature data from the l-1 layer to the l+1 layer is as follows:

[0043]

[0044] The improved Swin Transformer is connected in parallel with the CSP Bottleneck structure in the C2f module to form the C2f-ST feature extraction module. The powerful feature extraction capability of Swin Transformer is utilized to enable the backbone network to better extract the feature information of the input image.

[0045] (2) The neck network is responsible for fusing the extracted feature maps. It introduces the RFA (Residual Feature Augmentation) residual feature enhancement module and adopts the residual branch to improve the original FPN (Feature Pyramid Networks) structure in YOLOv8. It injects contextual information from different spaces into it to improve the feature expression of high-level features and enhance the network's ability to detect small targets. It also inserts a lightweight attention mechanism ECA module into the FPN structure to enhance the network's ability to focus on important features while having little impact on the network parameters and computational complexity.

[0046] ①RFA residual feature enhancement module: The specific construction steps of the RFA module are as follows: First, by performing a scale-invariant adaptive pooling operation on the feature S = h × w map of scale, different scales and depths are generated Multiple context features; each context feature is then convolved through a 1×1 convolution to reduce the channel feature dimension; finally, they are upsampled to the S scale by using bilinear interpolation for subsequent fusion operations. Considering the aliasing effect that may be caused by interpolation, an Adaptive Spatial Fusion (ASF) module is used to adaptively combine context features instead of simply using a summation operation. The ASF module takes the upsampled features as input and generates a spatial weight map for each feature. The context features are aggregated into M using weights and given context information at multiple scales.

[0047] ②ECA lightweight attention mechanism: The implementation of the ECA attention mechanism mainly consists of three steps: compression, excitation, and feature recalibration. First, the compression operation is the global average pooling operation (GAP), which compresses the input feature map into a 1×1C (C is the number of channels) real number sequence with global information in the spatial dimension to solve the problem of low information utilization between network layers. Then, the excitation operation corresponds to Fex(·,W) in the figure, which is also the difference between the ECA attention mechanism and the SE attention mechanism. The SE attention mechanism uses two fully connected layers to reduce the number of feature channels C to the original 1 / r, reducing the amount of calculation. The ECA attention mechanism abandons the intermediate dimensionality reduction operation and chooses to use each channel and its K nearest neighbors to obtain local cross-channel information, which not only avoids the information loss caused by dimensionality reduction, but also effectively captures the cross-channel information interaction. After the channel is restored to C, the weight value of each channel is obtained through the activation function. Finally, the feature recalibration step multiplies the weight value obtained in the compression step with the original feature channel by channel, thereby realizing the recalibration of the feature.

[0048] (3) Add a small target detection layer to the original neck network, and add a detection head for processing large-size feature maps to the head network. The specific method is: add the feature map of the second layer of the backbone part to the feature fusion network to help the network better capture shallow semantic information, thereby improving the detection accuracy of extremely small targets, as shown in the following formula:

[0049] P_2=C3(Concat(Upsample(P2 / 4_neck),P2 / 4_backbone))

[0050] Among them, P2 / 4neck means that the size of the P2 layer in Neck is reduced by 4 times relative to the size of the input image. Similarly, P2 / 4backbone means that the size of the P2 layer in the backbone network is reduced by 4 times relative to the size of the input image. Specifically, after two Upsample upsampling and Concat, another upsampling operation is performed to obtain a feature map of size 160×160. After upsampling and fusing features of different scales, the feature map has stronger semantic information and finer spatial information, so that it can better distinguish and detect very small targets. Then, the obtained feature map is concat-operated with the output of the second layer of the backbone network to further fuse feature information of different scales, and a C3 module is introduced to process the fused feature map, keeping its size at 160×160, and finally a very small target detection layer is obtained.

[0051] (4) The Dynamic Head detection head is introduced into the original head network. The attention mechanism unifies the detection heads of various targets. By utilizing the attention between feature levels of scale perception, the attention between spatial positions of spatial perception, and the attention between output channels of task perception, the expressiveness of the model target can be greatly enhanced, making the detection of small targets more flexible and accurate.

[0052] The scale-aware attention module fuses features of different scales according to their semantic importance as shown in the following equation.

[0053]

[0054] f(·) is a linear function approximated by a 1×1 convolution σ(x)=max(0,min(1,(x+1) / 2)).

[0055] The spatially aware attention module first uses deformable convolution to learn sparsity and then aggregates cross-layer features at the same spatial location. Where K is the number of sparsely sampled locations, p k +Δp k is the self-learned spatial shift focused on the shifted position of the unique region, Δm k is self-learning at position p k are important scalars at , and are all learned from the input features of the intermediate layer F.

[0056]

[0057] π C (F)·F=max(α 1 (F)·F C +β 1 (F),α 2 (F)·F C +β 2 (F)

[0058] The task-aware attention module dynamically opens or closes feature channels and selects different tasks, where θ(·) = [α 1 ,β 1 ,α 2 ,β 2 ] T , is a hyperfunction that learns to control the activation threshold. It first performs global pooling to reduce the dimension in the L×S dimension, then uses two fully connected layers, a normalization layer, and finally uses a shifted sigmoid function to normalize the output to [-1,1].

[0059] Specific implementation method of the present invention: The improved YOLOv8 target detection model is mainly used to perform feature extraction, feature fusion, generate prediction boxes and predict categories on scene images, and use the trained model to detect targets and coordinate boxes in scene images.

[0060] First, in the field scene, the drone equipment is used to collect the image data containing the target category from the top-down perspective, and the data set containing sensitive targets is collected on the Internet. The image data obtained by the two methods are merged to obtain the image set Images. The image set is annotated with the LabelImage software. The actual anchor box information of each target is saved in the yolo format (index, cx, cy, w, h). Index represents the index of the label name in the label array, cx and cy represent the x and y coordinates of the center point of the marked box, and w and h represent the width and height of the marked box. The values ​​are the result of dividing the width and height of the image. Each image corresponds to a label file, and each line of data in the label file corresponds to the anchor box information of a target contained in this image. Label all the collected image data, store the label file in the label set Labels, and put the image dataset Images and the label information dataset Labels into the dataset Final_Data. The data set is divided into the test set Test, the training set Train, and the validation set Val in the ratio of test set: training set: validation set = 1:8:1. The training set and validation set are sent to the model for training. During training, the modified model network file is selected. Thus, an improved detection model is trained using the scene data. The trained model is used to detect the images in the test set Test. The detection results are also output in the yolo format. This completes the generation of the sensitive target detection model under depression angle.

[0061] The generated model can be deployed on monitoring equipment or drone equipment to achieve real-time detection of sensitive targets in complex environments. It can improve the detection speed and accuracy of sensitive targets such as people, vehicles, tanks, warships, radars, etc. in scene images, and then obtain sensitive target category information and coordinate information, providing valuable auxiliary intelligence for emergency tasks.

[0062] The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A method for long-distance recognition of small targets with sensitive depression angle based on deep learning. It is characterized in that Its basic network YOLOv8 target detection model consists of three parts: backbone network, neck network and head network. Firstly, the improved Swin Transformer structure is introduced into the original backbone network to enhance the backbone network's feature extraction ability for input images. Then, the original FPN structure is replaced with a lighter and more efficient residual feature fusion pyramid structure RE-FP N, and a small target detection layer is added to improve the network's detection ability for objects of different scales, especially small targets; A dynamic detection head Dynamic Head with a multiple attention mechanism is introduced to enable the model to pay more attention to dense small target areas and extract more small target features to improve the detection speed and accuracy of sensitive targets such as people, vehicles, and important infrastructure in scene images, and then obtain sensitive target category information and coordinate information, providing valuable auxiliary intelligence for complex emergency tasks.

2. According to the method for long-distance recognition of small targets sensitive to depression angle based on deep learning in claim 1, It is characterized in that The backbone network is used to perform feature extraction operations on the input model image. The improved Swin Transformer structure is introduced into the C2f feature extraction module in the original backbone network to form a C2f-ST feature extraction module, and the C2f module is used to replace all the C2f modules of the original backbone network to obtain a new feature extraction backbone network, enhance the feature extraction capability of the network, better retain context information, so as to improve the detection accuracy of small targets in the input image.

3. According to claim 1, the method for long-distance recognition of small targets sensitive to depression angle based on deep learning, It is characterized in that The neck network is responsible for fusing the extracted feature maps, introducing the RFA (Residual Feature Augmentation) residual feature enhancement module, using residual branches, improving the original FPN (Feature Pyramid Networks) structure in YOLOv8, injecting contextual information from different spaces into it to improve the feature expression of high-level features, and enhancing the network's ability to detect small targets. A lightweight attention mechanism ECA module is inserted into the FPN structure to improve the network's ability to focus on important features with little impact on the network parameters and computational complexity.

4. According to claim 1, the method for long-distance recognition of small targets sensitive to depression angle based on deep learning, Features A small target detection layer is added to the original neck network, and a detection head for processing large-size feature maps is added to the head network. In this way, a set of 160x160 feature maps to be detected is added on the basis of the original 20x20, 40x40, and 80x80 features to be detected, which helps the network better capture shallow semantic information and uses multi-scale features to help the model better handle targets of different sizes.

5. The method for long-distance recognition of depression-sensitive small targets based on deep learning according to claim 1, It is characterized in that The head network complexly processes the detection feature map and outputs sequence vectors of detection results of different scales. We use Dynamic Head to replace the original detection head and introduce a target detection head method with multiple attention mechanisms to further process the output feature map. We introduce three attention mechanisms in different dimensions to improve its adaptability in scale perception, spatial position and multi-task, significantly improving the expression ability of the detection head.

6. The method for long-distance recognition of depression-sensitive small targets based on deep learning according to claim 2, It is characterized in that The Swin Transformer module is composed of a layer normalization module (LN), a window-based multi-head self-attention module (W-MSA), a shifted window-based multi-head self-attention module (SW-MSA), and a two-layer multi-layer perceptron (MLP) module containing a GELU nonlinear layer. The specific process of input feature data from the l-1 layer to the l+1 layer is shown as follows: LN can make the network better integrated and prevent the network from overfitting. However, using LN layers in CNN will destroy the learned sample features. Therefore, in order to preserve the generalization ability of the model while protecting CNN's ability to learn sample features, the LN in the input part of the cascaded Swin Transformer is removed to make it more suitable for the YOLOv8 network. The specific process of improving the input feature data from the l-1 layer to the l+1 layer is as follows: The improved Swin Transformer is connected in parallel with the CSP Bottleneck structure in the C2f module to form the C2f-ST feature extraction module. The powerful feature extraction capability of Swin Transformer is utilized to enable the backbone network to better extract the feature information of the input image.

7. The method for long-distance recognition of small targets sensitive to depression angles based on deep learning according to claim 3, It is characterized in that The RFA (Residual Feature Augmentation) residual feature enhancement module reduces the spatial information loss of the feature map extracted by the backbone network by collecting spatial context information, and at the same time improves the performance of the feature pyramid. Considering the aliasing effect that may be caused by interpolation, the ASF (Adaptive Spatial Fusion) adaptive spatial fusion module is introduced to adaptively combine context features, generate a spatial weight map for each feature, and use the weights to aggregate the context features into M, giving it context information at multiple scales.

8. The method for long-distance recognition of depression-sensitive small targets based on deep learning according to claim 3, It is characterized in that The ECA (Efficient Channel Attention) attention mechanism reconstructs the calculation weights of different parts of the feature map during feature fusion, so that the model can improve its attention to the target area. The ECA attention mechanism consists of three steps: compression, excitation and feature recalibration. It greatly reduces the amount of network parameters while avoiding the side effects of dimensionality reduction operations.

9. The method for long-distance recognition of depression-sensitive small targets based on deep learning according to claim 5, It is characterized in that The Dynamic Head dynamic detection head and attention mechanism unify the detection heads of various targets. By utilizing the attention between feature levels of scale perception, the attention between spatial positions of spatial perception, and the attention between output channels of task perception, the expressiveness of the model target can be greatly enhanced, and the detection of small targets can be more flexible and accurate. The scale-aware attention module fuses features of different scales according to their semantic importance as shown in the following equation. f(·) is a linear function σ(x)=max(0, min(1, (x+1) / 2)) approximated using a 1×1 convolution. The spatially aware attention module first uses deformable convolution to learn sparsity and then aggregates cross-layer features at the same spatial location. Where K is the number of sparsely sampled locations, p k +Δp k is the self-learned spatial shift focused on the shifted position of the unique region, Δm k is self-learning at position p k are important scalars at , and are all learned from the input features of the intermediate layer F. p C (F)·F=max(α 1 (F)·F C +b 1 (F),a 2 (F)·F C +b 2 (F)) The task-aware attention module dynamically opens or closes feature channels and selects different tasks, where θ(·) = [α 1 , β 1 , α 2 , β 2 ] T , which is a hyperfunction that learns to control the activation threshold, first performs global pooling to reduce the dimension in the L×S dimension, then uses two fully connected layers, a normalization layer, and finally uses a shifted sigmoid function to normalize the output to [-1, 1].

Images