A method for detecting small aerial targets based on deep learning

Abstract

The present invention discloses a method for detecting small aerial targets based on deep learning, comprising: obtaining an aerial image to be detected; inputting the aerial image to be detected into a target detection model to obtain an initial detection result; wherein the target detection model is trained using a training set comprising aerial images and annotations of extreme environments; the target detection model is constructed using a neural network, the neural network is used to perform multiple image fusions on the aerial image to be detected to extract a confidence map, and detection results are obtained based on the confidence map. The method of the present invention introduces a feature embedding module into the small aerial target detection model to assist in detection, enabling excellent detection results even in complex backgrounds and low signal-to-noise ratio conditions.

Description

Technical Field

[0001] The present invention belongs to the technical field of target detection, and in particular relates to a method for detecting small aerial targets based on deep learning. Background Art

[0002] Aerial target detection technology is particularly important in complex backgrounds and low-contrast conditions. This is especially true in applications such as drone surveillance, aerial reconnaissance, and emergency response. Low contrast and environmental interference often blur or lose target features, compromising detection and recognition accuracy. For example, in drone surveillance, environmental noise can hinder the detection of potential targets; in aerial reconnaissance, reduced detection system effectiveness in low-light conditions can compromise mission success. Therefore, improving the robustness of aerial target detection technology is crucial for improving the accuracy of these tasks in harsh environments.

[0003] Existing methods for detecting small and dim targets in aerial imagery primarily include algorithms based on physical models and deep learning. Physical-based methods, such as image enhancement and denoising, improve target detection by increasing the overall contrast or brightness of specific areas in aerial images, but their performance is limited in low signal-to-noise ratio conditions. Deep learning methods, such as YOLO and Faster R-CNN, learn target detection and recognition features by training neural networks on large-scale datasets. These methods demonstrate good detection performance in complex backgrounds, but still have limitations in extreme environments, such as those with high noise and low contrast. Summary of the Invention

[0004] To solve the above technical problems, the present invention proposes a deep learning-based method for detecting small aerial targets, which can demonstrate good detection performance in extreme environments.

[0005] The present invention provides a method for detecting small aerial targets based on deep learning, comprising:

[0006] Acquire the aerial image to be detected;

[0007] The aerial image to be detected is input into a target detection model to obtain an initial detection result; wherein, the target detection model is obtained by training a training set, and the training set includes: aerial images of extreme environments and annotations. The target detection model is constructed using a neural network, and the neural network is used to perform multiple image fusion on the aerial image to be detected to extract a confidence map, and obtain the detection result based on the confidence map.

[0008] Optionally, the target detection model includes: a feature embedding module, a position encoding module, a composite encoding module, an upsampling module and a decoder;

[0009] The feature embedding module is used to convert the low-level feature representation of the input image into a high-dimensional feature representation;

[0010] The position encoding module is used to add spatial position information to the input sequence;

[0011] The composite encoding module is used to extract local and global features and provide multi-scale semantic information;

[0012] The upsampling module is used to obtain a high-resolution feature map;

[0013] The decoder is used to decode the high-resolution feature map to obtain an initial detection result.

[0014] Optionally, the feature embedding module includes: a convolutional layer, an activation layer, and a plurality of convolutional blocks connected in sequence;

[0015] The convolution block includes: a convolution layer, a batch normalization layer and an activation layer connected in sequence.

[0016] Optionally, the position encoding module includes: a feature embedding layer and an encoding layer;

[0017] The feature embedding layer is used to segment and linearly embed the input image to generate an embedded feature sequence;

[0018] The coding layer is used to introduce spatial information to perform position coding on the feature sequence to obtain a sequence feature graph.

[0019] Optionally, the composite encoding module includes: a multi-head self-attention unit and an enhancement unit;

[0020] The multi-head self-attention unit is used to obtain global context information based on the sequence feature graph;

[0021] The enhancement unit is used to enhance feature expression by aggregating context information at different positions and generate an enhanced feature map with the same dimension as the sequence feature map.

[0022] Optionally, the multi-head self-attention unit includes: a self-attention mechanism layer and a global normalization layer;

[0023] The enhancement unit includes: a first convolutional layer and a second convolutional layer.

[0024] Optionally, the step of obtaining global context information includes:

[0025] Inputting the sequence feature graph into the self-attention mechanism layer in sequence to obtain a first result graph;

[0026] Inputting the sequence feature graph into a global normalization layer to obtain a second result graph;

[0027] The first result graph and the second result graph are fused to obtain the global context information.

[0028] Optionally, generating an enhanced feature map having the same dimension as the sequence feature map includes:

[0029] Inputting the global context information into the first convolutional layer and the second convolutional layer in sequence to obtain a third result graph;

[0030] The third result map and the global context information are weightedly fused to obtain the enhanced feature map.

[0031] Optionally, decoding the high-resolution feature map to obtain a detection result includes:

[0032] The high-resolution feature map is fused with the sequence feature map, and the fused feature map is decoded to obtain the detection result.

[0033] Optionally, the loss function is:

[0034]

[0035] Among them, L iou is the loss function, y is the confidence map, x gt is a real labeled image.

[0036] Compared with the prior art, the present invention has the following advantages and technical effects:

[0037] The method of the present invention introduces a feature embedding module into the aerial small target detection model to assist detection, so that excellent detection results can be achieved even under complex backgrounds and low signal-to-noise ratio conditions. The model uses aerial image features as prior information related to the characteristics of small targets and introduces a feature extraction strategy that combines a Transformer module with a composite encoder module. The Transformer module can extract global context information of the target area, which complements the local feature information of the composite encoder module to further improve detection performance. In addition, a multi-head self-attention mechanism and a layer normalization module are designed in the context enhancement module of the detection model to better identify key feature channels and improve the detection accuracy of the model. The context enhancement module can effectively focus on important areas in the image and reduce attention to irrelevant background, so that the model can achieve efficient small target detection in different environments, ultimately improving detection performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. In the accompanying drawings:

[0039] Figure 1 This is a flow chart of a method for detecting small aerial targets based on deep learning according to an embodiment of the present invention;

[0040] Figure 2 1 is a schematic structural diagram of image detection according to an embodiment of the present invention;

[0041] Figure 3 Schematic diagram of the local structure of the target detection model according to an embodiment of the present invention;

[0042] Figure 4 is a schematic diagram of a Transformer module according to an embodiment of the present invention;

[0043] Figure 5 This is a diagram showing the effect of detecting small and weak targets in the air according to an embodiment of the present invention. DETAILED DESCRIPTION

[0044] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0045] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0046] This paper proposes a method for detecting small and weak aerial targets based on deep learning. Figure 1 As shown, the specific steps include:

[0047] Acquire the aerial image to be detected;

[0048] The aerial image to be detected is input into the target detection model to obtain the initial detection result; wherein, the target detection model is obtained by training a training set, which includes aerial images of extreme environments and annotations. The target detection model is constructed using a neural network, which is used to perform multiple image fusion on the aerial image to be detected to extract the confidence map, and obtain the detection result based on the confidence map.

[0049] Furthermore, the object detection model includes: a feature embedding module, a position encoding module, a composite encoding module, an upsampling module and a decoder;

[0050] Feature embedding module, used to transform the low-level feature representation of the input image into a high-dimensional feature representation;

[0051] Position encoding module, used to add spatial position information to the input sequence;

[0052] Composite encoding module, used to extract local and global features and provide multi-scale semantic information;

[0053] Upsampling module, used to obtain high-resolution feature maps;

[0054] The decoder is used to decode the high-resolution feature map and obtain the initial detection results.

[0055] Specifically, such as Figure 2 As shown, obtain the image to be detected;

[0056] The aerial image is input into the trained object detection model, in which the aerial image is passed through the feature embedding module to obtain the first feature map;

[0057] The first feature map is segmented and linearly embedded to generate an embedded feature sequence, and position encoding is performed to obtain a second feature map;

[0058] The second feature map is input into the composite encoder module, which includes a preset number of feature extraction structures connected in sequence. When passing through the feature extraction structure, the input of each feature extraction structure passes through the multi-head self-attention module (MSA), the layer normalization module, the context enhancement module (CAM) and the first activation layer in sequence to obtain a third feature map, and then the third feature map is fused with the input of the feature extraction structure to obtain a fourth feature map; the fourth feature map then passes through the second multi-head self-attention module, the second layer normalization module, the context enhancement module (CAM) and the second activation layer in the feature extraction structure in sequence to generate a fifth feature map, and then the fifth feature map is fused with the fourth feature map to obtain a sixth feature map, and the sixth feature map is fused with the input of the feature extraction structure to obtain a seventh feature map as the output feature of the composite encoder module.

[0059] The seventh feature map is processed through an upsampling module to obtain a high-resolution eighth feature map;

[0060] In the decoding module, the eighth feature map is jump-connected to the second feature map to obtain a ninth feature map;

[0061] Decode the ninth feature map to generate a confidence map.

[0062] Finally, the confidence map is input into the output module to obtain the detection result map.

[0063] Furthermore, the feature embedding module includes: a convolutional layer, an activation layer, and several convolutional blocks connected in sequence;

[0064] The convolution block consists of: a convolutional layer, a batch normalization layer, and an activation layer connected in sequence.

[0065] Specifically, the feature embedding module includes a convolution layer, an activation layer, and several convolution blocks connected in sequence from data input to output. Each convolution block includes a convolution layer, a batch normalization layer, and an activation layer connected in sequence from data input to output.

[0066] Furthermore, the position encoding module includes: a feature embedding layer and an encoding layer;

[0067] The feature embedding layer is used to segment and linearly embed the input image to generate an embedded feature sequence;

[0068] The encoding layer is used to introduce spatial information to perform position encoding on the feature sequence and obtain the sequence feature map.

[0069] Specifically, the encoding module contains a feature embedding layer, which performs block segmentation and linear embedding on the input aerial image to generate an embedded feature sequence, and introduces spatial information through position encoding to obtain a sequence feature map.

[0070] Furthermore, the composite encoding module includes: a multi-head self-attention unit and an enhancement unit;

[0071] Multi-head self-attention unit, used to obtain global context information based on sequence feature graph;

[0072] The enhancement unit is used to enhance feature expression by aggregating contextual information at different positions and generate an enhanced feature map with the same dimension as the sequence feature map.

[0073] Specifically, the composite encoder module combines the multi-head self-attention module (MSA) and the context enhancement module (CAM), and performs feature extraction through multiple stages, each of which includes a multi-head self-attention layer and a context enhancement layer connected in sequence.

[0074] The multi-head self-attention module (MSA) in the composite encoder module is used to calculate the global correlation of input features and extract global context information; the context enhancement module (CAM) enhances feature expression by aggregating context information at different positions and generates an enhanced feature map with the same dimension as the input feature map.

[0075] Furthermore, the multi-head self-attention unit includes: a self-attention mechanism layer and a global normalization layer;

[0076] The enhancement unit includes: a first convolutional layer and a second convolutional layer.

[0077] Furthermore, it is characterized in that obtaining global context information includes:

[0078] Input the sequence feature graph into the self-attention mechanism layer in sequence to obtain the first result graph;

[0079] Input the sequence feature graph into the global normalization layer to obtain a second result graph;

[0080] The first result graph and the second result graph are fused to obtain global context information.

[0081] Specifically, the multi-head self-attention module includes a parallel self-attention mechanism and a global normalization module. The output of the second-layer normalization module is processed by the self-attention mechanism and the global normalization module respectively. Then, the output of the self-attention mechanism is fused with the output of the global normalization module and input into the context enhancement module (CAM).

[0082] Furthermore, generating an enhanced feature map of the same dimension as the sequence feature map includes:

[0083] Input the global context information into the first convolutional layer and the second convolutional layer in sequence to obtain a third result graph;

[0084] The third result map and the global context information are weightedly fused to obtain an enhanced feature map.

[0085] Specifically, the context enhancement module includes a first convolutional layer and a second convolutional layer. The output of the second multi-head self-attention module is processed by the first convolutional layer and the second convolutional layer in sequence, and then fused with the output of the second multi-head self-attention module to obtain an intermediate feature map. At the same time, the output of the second multi-head self-attention module is processed by the second convolutional layer in sequence, fused with the intermediate feature map, and then input into the first activation layer.

[0086] Furthermore, the high-resolution feature map is decoded to obtain the detection results, including:

[0087] The high-resolution feature map is fused with the sequence feature map, and the fused feature map is decoded to obtain the detection result.

[0088] Specifically, the decoding module includes a preset number of upsampling layers connected in sequence, each of which includes a transposed convolution layer for restoring the high-dimensional feature map output by the composite encoder to the spatial resolution of the target detection result.

[0089] The upsampling module is used to restore the spatial resolution of the feature map and perform jump connections with the output features of each stage of the composite encoder, so that low-level detail information is integrated with high-level semantic information to improve the accuracy of target detection; the final output feature map is decoded to generate a confidence map and a detection result map.

[0090] Furthermore, the loss function is:

[0091]

[0092] Among them, L iouis the loss function, y is the confidence map, x gt is a real labeled image.

[0093] Specifically, the loss function in the training of the small aerial target detection model uses the intersection over union (IoU) to calculate the distance between the confidence map and the ground truth annotation, helping to address the class imbalance between the target and the background, and focusing more on the area of ​​small aerial targets. It is defined as follows:

[0094]

[0095] Among them, y is the confidence map, x gt is a real labeled image.

[0096] Furthermore, the preconditions are:

[0097] The initial detection result partially overlaps with the true annotation and the pixel distance between the center of the initial detection result and the center of the true annotation is less than a preset threshold, where the center of the detection result is obtained by calculating the centroid of the binary segmentation mask.

[0098] Specifically, the conditions for the detection result to be correct are that the following two points are met at the same time: (1) the output result overlaps with the true annotation in some pixels; (2) the pixel distance between the center of the detection result and the center of the true annotation is less than a certain threshold (4 pixels). Five widely used evaluation indicators are used, including detection probability (P d ), P d -F a The area under the curve (AUC), the target level F1 value (F t 1) Pixel level (F p 1) and Intersection over Union (IoU), these indices are defined as follows:

[0099]

[0100] Target level F1 value F t 1 and pixel level F1 value F p 1 is defined as follows:

[0101]

[0102] The precision and recall are defined as:

[0103]

[0104] Among them, T p is a true example, F p It is a false true example, F N is a false negative example. The Intersection over Union (IoU) is a pixel-level metric defined as:

[0105]

[0106] The present embodiment will be described in detail below with reference to the accompanying drawings:

[0107] like Figure 1-Figure 5 As shown, in one embodiment, a method for detecting small aerial targets based on deep learning includes:

[0108] Acquire the image to be detected;

[0109] The image is input into the trained small object detection model, in which:

[0110] The aerial image is processed through the feature embedding module to obtain the first feature map, and then the embedded feature sequence is generated through block segmentation and linear embedding, and position encoding is performed to obtain the second feature map;

[0111] Among them, the feature embedding module includes a convolution layer, an activation layer and multiple convolution blocks connected in sequence from data input to output, and each convolution block includes a convolution layer, a batch normalization layer and an activation layer connected in sequence from data input to output.

[0112] It should be noted that after the aerial image is input into the target detection model, the number of channels will be increased when passing through the feature embedding module to extract the features of the aerial image. The size of the input aerial image is C×H×W (where C represents the number of channels, H and W represent the height and width of the image respectively, and the specific size is determined according to the actual input image). It passes through the convolution layer (using a 3×3 convolution layer, at this time the number of channels increases from 3 to 64), the activation layer (using the softmax activation function, and the feature map size obtained at this time is 64×H×W) and multiple sequentially connected convolution blocks (the first feature map is obtained at this time, and the size of the first feature map is 64×H×W), where each convolution block includes a convolution layer (using a 3×3 convolution layer), a batch normalization layer and an activation layer (using a softmax activation function) connected in sequence from data input to output.

[0113] Positional encoding helps the model understand the relative or absolute position of each element in the input data when processing sequence data. Positional encoding adds a positional encoding vector to the input embedding vector to provide the position information of each element in the sequence. Specifically, for the posth position in the input sequence and the embedding dimension i:

[0114] When i is an even number:

[0115] When i is an odd number:

[0116] Calculate the encoding vector for each position to obtain a matrix with the same length as the input sequence, and add this matrix to the input embedding matrix element by element to obtain the final input representation.

[0117] The second feature map is input into the composite encoder module to obtain a third feature map;

[0118] Among them, such as Figure 4 As shown in the figure, the composite encoder module is based on the Transformer model and includes multiple feature extraction structures in sequence. Each feature extraction structure contains a multi-head self-attention module (MSA), a layer normalization module, a context enhancement module (CAM), and the first activation layer. After the second feature map (size 64xHxW) is input, it passes through the first feature extraction structure. After passing through the multi-head self-attention module, the number of channels of the feature map increases to 128, while the height and width remain unchanged. The resulting third feature map size is 128xHxW. Next, after layer normalization and activation function processing, the feature map still maintains this size, but the eigenvalue distribution is optimized.

[0119] The third feature map is fused with the second feature map to obtain a fourth feature map;

[0120] During the fusion process, the third feature map is added to the input second feature map (size 64xHxW), and the size of the obtained fourth feature map remains 128xHxW, combining feature information from two different sources.

[0121] The fourth feature map is processed by the second feature extraction structure in turn to obtain the fifth feature map;

[0122] After the second feature extraction structure, the feature map passes through the multi-head self-attention module again, and its size becomes: the height and width are halved, and the number of channels is doubled. It is then processed through layer normalization and a second activation layer to further enhance the feature expression capability.

[0123] Fusing the fifth feature map with the fourth feature map to obtain a sixth feature map;

[0124] Among them, the fifth feature map after processing by the second feature extraction structure has a size of Add the fourth feature map to obtain the sixth feature map size To further integrate feature information at different levels.

[0125] The sixth feature map is then passed through the attention module, residual module, decoding module, and output module in sequence to finally obtain the seventh feature map. (The backbone module, encoding module, attention module, residual module, decoding module, and output module constitute the CNN network designed by the present invention).

[0126] The seventh feature map is input into the upsampling module, and the eighth feature map is obtained after upsampling;

[0127] The upsampling module uses bilinear interpolation to upsample the feature map to 256xHxW. The upsampling module interpolates the pixels of each feature map, expanding the spatial dimension of the feature map for effective feature fusion with the subsequent decoding module.

[0128] In the decoding module, the eighth feature map is jump-connected to the second feature map to obtain a ninth feature map;

[0129] The second feature map has a size of 64xHxW. By concatenating it with the eighth feature map (this concatenation is channel-wise), the resulting ninth feature map has a size of 320xHxW (256 + 64 channels), enhancing the model's ability to recover detailed information. This skip connection ensures that important low-level features are passed to high-level structures during decoding, thereby improving the performance of the final output.

[0130] Subsequently, the ninth feature map passes through a series of convolutional layers, activation functions, and the final output layer in the decoding module to generate the final output feature map;

[0131] The convolutional layers in the decoding module use a 3 x 3 convolution kernel, extracting deep features and gradually restoring spatial resolution through layer-by-layer convolution. After the first convolutional layer, the ninth feature map has a size of 320 x H x W. The activation layer (using the ReLU activation function) maintains the nonlinearity of the feature map. Subsequently, after the second convolutional layer and the activation layer, the resulting feature map still has a size of 320 x H x W.

[0132] Finally, after passing through the output layer (usually a 1x1 convolutional layer), the output feature map is obtained. The output feature map is as follows Figure 5 As shown in the figure, the size is NxHxW, where N is the number of categories output by the model. This process helps the model make full use of contextual information and effectively detect small and weak targets in the air, improving detection precision and accuracy.

[0133] The dataset used for model training in this embodiment consists of 15,064 images of various flying objects, with a training rate of 80% and a verification rate of 20%. Each image is annotated with the class number of the object and the coordinates of the edge of the relevant bounding box. An image may have multiple objects and classes, with an average of 1.6 annotated objects per image and a total of 24,769 annotations for all images. The median image size is 416x416. These images have been pre-processed for automatic orientation, but no enhancement has been applied. The dataset represents a long-tail distribution, with drones (25.2% of objects), birds (25%), airplanes (7.9%), and helicopters (6.3%) accounting for the majority of the dataset (64.4%), and there is a class imbalance.

[0134] The framework of the method proposed in this embodiment is implemented using Pytorch 1.7.1 and accelerated by CUDA 11.2. The entire network is trained using the SGD algorithm with a learning rate of 0.01, a momentum of 0.9, and a weight decay of 1e-4, running on an NVIDIA GeForce RTX4050 GPU. The batch size is set to 24 and the maximum number of training rounds is 150. Before training, all input images are normalized. Subsequently, the images are data augmented by random flipping. Finally, the images are resized to 224x224. The number of heads of MSA is set to 12, and the number of encoding layers is set to 12.

[0135] The loss function used in the training of the small aerial target detection model uses the intersection over union (IoU) to calculate the distance between the confidence map and the ground truth annotation. This helps address the class imbalance between the target and the background, and focuses more on the area of ​​small aerial targets. It is defined as follows:

[0136]

[0137] Among them, y is the confidence map, x gt is a real labeled image.

[0138] The method for detecting small aerial targets based on deep learning is characterized in that the correct detection result is when the following two conditions are met simultaneously: (1) the output result overlaps with the true annotation in some pixels; (2) the pixel distance between the center of the detection result and the center of the true annotation is less than a certain threshold (4 pixels). Five widely used evaluation indicators are used, including detection probability (P d ), P d -F a The area under the curve (AUC), the target level F1 value (F t 1) Pixel level (F p 1) and Intersection over Union (IoU). These indices are defined as follows:

[0139]

[0140] Target level F1 value F t 1 and pixel level F1 value F p 1 is defined as follows:

[0141]

[0142] The precision and recall are defined as:

[0143]

[0144] Among them, T pis a true example, F p It is a false true example, F N is a false negative example. The Intersection over Union (IoU) is a pixel-level metric defined as:

[0145]

[0146] The deep learning-based method for detecting small aerial targets introduces a feature embedding module into the small aerial target detection model to assist in detection, so that it can also show excellent detection results under complex backgrounds and low signal-to-noise ratio conditions. At the same time, the aerial image is regarded as prior information related to the target features, and the Transformer module is combined with the improved UNet network to extract features. The Transformer can obtain global information containing target features, which complements the local feature information extracted by the UNet, thereby improving detection performance; the target detection model uses attention to help the model better focus on the key feature channels in the image, thereby improving the detection accuracy of the model. The attention mechanism enables the model to focus on important areas in the image, reduce interference with the background, and enable the model to accurately detect small targets even in complex backgrounds, thereby improving detection performance.

[0147] The above are merely preferred embodiments of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A method for detecting small aerial targets based on deep learning, characterized in that: include: Acquire the aerial image to be detected; Inputting the aerial image to be detected into a target detection model to obtain an initial detection result; wherein the target detection model is obtained by training a training set, wherein the training set includes: aerial images of extreme environments and annotations; the target detection model is constructed using a neural network, and the neural network is used to perform multiple image fusion on the aerial image to be detected to extract a confidence map, and obtain a detection result based on the confidence map; The target detection model includes: a feature embedding module, a position encoding module, a composite encoding module, an upsampling module and a decoder, wherein the second feature map is obtained after being processed by the feature embedding module and the position encoding module, and the composite encoding module further includes an enhancement unit; The second feature map is input into a composite coding module, which includes a preset number of feature extraction structures connected in sequence. When the input of each feature extraction structure passes through the feature extraction structure, it passes through the multi-head self-attention unit, the layer normalization module, the enhancement unit and the first activation layer in sequence to obtain a third feature map, and then the third feature map is fused with the input of the feature extraction structure to obtain a fourth feature map; the fourth feature map is then passed through the second multi-head self-attention unit, the second layer normalization module, the enhancement unit and the second activation layer in the feature extraction structure in sequence to generate a fifth feature map, and then the fifth feature map is fused with the fourth feature map to obtain a sixth feature map, and the sixth feature map is fused with the input of the feature extraction structure to obtain a seventh feature map as the output feature of the composite coding module; The enhancement unit includes a first convolutional layer and a second convolutional layer. The output of the second multi-head self-attention unit is processed by the first and second convolutional layers in sequence, and then fused with the output of the second multi-head self-attention unit to obtain an intermediate feature map. At the same time, the output of the second multi-head self-attention unit is processed by the second convolutional layer in sequence, fused with the intermediate feature map, and then input into the first activation layer to obtain an enhanced feature map.

2. The method for detecting small aerial targets based on deep learning according to claim 1, wherein: The feature embedding module is used to convert the low-level feature representation of the input image into a high-dimensional feature representation; The position encoding module is used to add spatial position information to the input sequence; The composite encoding module is used to extract local and global features and provide multi-scale semantic information; The upsampling module is used to obtain a high-resolution feature map; The decoder is used to decode the high-resolution feature map to obtain an initial detection result.

3. The method for detecting small aerial targets based on deep learning according to claim 2, wherein: The feature embedding module includes: a convolutional layer, an activation layer and several convolutional blocks connected in sequence; The convolution block includes: a convolution layer, a batch normalization layer and an activation layer connected in sequence.

4. The method for detecting small aerial targets based on deep learning according to claim 2, wherein: The position encoding module includes: a feature embedding layer and an encoding layer; The feature embedding layer is used to segment and linearly embed the input image to generate an embedded feature sequence; The coding layer is used to introduce spatial information to perform position coding on the feature sequence to obtain a sequence feature graph.

5. The method for detecting small aerial targets based on deep learning according to claim 4, wherein: The composite coding module includes: a multi-head self-attention unit and an enhancement unit; The multi-head self-attention unit is used to obtain global context information based on the sequence feature graph; The enhancement unit is used to enhance feature expression by aggregating context information at different positions and generate an enhanced feature map with the same dimension as the sequence feature map.

6. The method for detecting small aerial targets based on deep learning according to claim 5, characterized in that: The multi-head self-attention unit includes: a self-attention mechanism layer and a global normalization layer.

7. The method for detecting small aerial targets based on deep learning according to claim 6, wherein: Obtaining global context information includes: Inputting the sequence feature graph into the self-attention mechanism layer in sequence to obtain a first result graph; Inputting the sequence feature graph into a global normalization layer to obtain a second result graph; The first result graph and the second result graph are fused to obtain the global context information.

8. The method for detecting small aerial targets based on deep learning according to claim 4, wherein: Decoding the high-resolution feature map to obtain a detection result includes: The high-resolution feature map is fused with the sequence feature map, and the fused feature map is decoded to obtain the detection result.

9. The method for detecting small aerial targets based on deep learning according to claim 4, wherein: The loss function is used for convergence during the training process, and the loss function is: Among them, L iou is the loss function, y is the confidence map, x gt is a real labeled image.

Images