A UAV target detection method based on lightweight feature fusion

Abstract

This paper discloses a drone target detection method based on lightweight feature fusion, which can identify targets in drone aerial images. To address the problems of low accuracy in detecting small targets in images and the large number of network parameters that make real-time detection difficult in current target detection networks, this paper first redesigns the YOLOv4‑tiny neck feature extraction module using depthwise separable convolution (DSC) and coordinate attention (CA). The SPPF feature extraction module is then used to increase the receptive field information of the extracted features while effectively maintaining the computational lightweight of the model. Finally, the Decoupled Head detection module is used to enhance the spatial association of the extracted feature information, making it easier to distinguish between the background and the target to be detected, thereby constructing a new lightweight target detection network.

Description

Technical Field

[0001] The present invention relates to the field of target detection technology, and more specifically, to a drone target detection method based on lightweight feature fusion. Background Art

[0002] With the continuous development of deep learning, object detection has become one of the hottest research areas. Object detection algorithms can be applied to scenarios of great significance in real life, such as transportation, industrial inspection, facial recognition, and military surveillance. Common object detection algorithms are generally divided into two-stage and single-stage algorithms. Two-stage detection algorithms first generate candidate regions and extract pre-selected bounding boxes within them. They then simultaneously classify and predict the detection category and the location of the target region. Their advantages include high average precision and low false positive and false negative rates. However, their large number of model parameters and computational complexity make them difficult to meet real-time requirements. Typical examples include Fast R-CNN and Faster R-CNN. Single-stage detection algorithms, such as the YOLO series and SSD algorithms, directly integrate feature extraction and predicted bounding box location. They offer advantages such as high detection speed, low model complexity, and ease of deployment, making them suitable for real-time detection.

[0003] Compared with the current general target detection algorithms, which have shortcomings such as small field of view, many lateral obstructions to the detection target, and difficulty in field of view conversion, drone target detection not only has the advantages of unobstructed lateral field of view at high altitude, wide monitoring range, and rapid field of view conversion, but can also be used for target detection in military and various dangerous scenarios, which can make the application of target detection algorithms safer. Therefore, target detection algorithms based on drone aerial images have become a research hotspot.

[0004] However, drone aerial images also have a series of problems such as small target size, complex background, and blurred image appearance, which lead to reduced detection accuracy. The most difficult problem to solve is the low accuracy of small target detection. In small target detection, network training and target prediction mainly focus on feature extraction and feature fusion. Since small targets occupy a small number of pixels in the original image, they carry limited information and lack appearance information such as texture, shape, and color. In addition, the downsampling caused by deep convolution causes the feature information of small targets to diffuse or even disappear in the deep feature map. Therefore, enriching the network's contextual semantic information, location information, and feature representation has become a research focus for drone small target detection tasks. At the same time, drone memory and computing resources are limited. Therefore, target detection algorithms based on drone aerial images must control the growth of model parameters and keep the model lightweight to maintain real-time detection speed.

[0005] Current deep learning-based drone target detection techniques typically integrate and improve classic deep learning target detection algorithms by combining drone technology and the characteristics of aerial imagery. Among target detection algorithms with high speed requirements, most incorporate the characteristics of drone detection to improve the YOLO family of algorithms. Due to the limitations of drone computing power, Zhang et al. pruned the original YOLOv3 network to obtain Slim YOLOv3, which maintains essentially the same accuracy but significantly reduces the number of parameters, memory usage, and inference time. This improved model is more suitable for drone deployment. Zhu et al. proposed TPH-YOLOv5, which replaces the original YOLO prediction heads with transformers based on YOLOv5, increasing the number of heads from three to four. Furthermore, the Convolutional Block Attention Module (CBAM) attention mechanism is integrated into the feature fusion stage, significantly improving detection accuracy.

[0006] However, most drone-based object detection algorithms focus primarily on improving either detection accuracy or speed, without properly balancing accuracy and speed. Therefore, designing an algorithm that effectively balances accuracy and speed can accelerate the development of drone applications in the field of object detection and is a crucial research topic for social development.

[0007] Drone images are often characterized by high scene complexity due to factors such as atmospheric temperature and lighting conditions. These factors include a wide variety of foreground objects, varying shapes, and scales, making target recognition difficult. Accurately detecting targets of interest in drone imagery is a challenging task. Furthermore, the limited computing resources and capabilities of drone-based equipment conflict with the requirements for real-time detection. Balancing detection accuracy and speed has always been a challenge for researchers and developers. Balancing the accuracy and speed of deep learning-based target detection models within the constraints of hardware resources remains a core research issue. However, most current drone-based target detection algorithms focus primarily on improving either accuracy or speed, failing to achieve a good balance between accuracy and speed. Summary of the Invention

[0008] The present invention is provided to solve the above-mentioned problems existing in the prior art. The present invention provides a UAV target detection method based on lightweight feature fusion, which can perform target recognition on images taken by UAVs. To address the problems of low detection accuracy of small targets in images and large number of network parameters that make it difficult to detect in real time in the current target detection network, the neck feature extraction module of YOLOv4-tiny is first redesigned by using depthwise separable convolution (DSC) and coordinate attention (CA); then the SPPF feature extraction module is used to increase the receptive field information of the extracted features and effectively keep the computation of the model lightweight; finally, the Decoupled Head detection head module is used to make the feature information finally extracted have stronger spatial union, making it easier to distinguish between the background and the target to be detected, thereby constructing a new lightweight target detection network.

[0009] The present invention specifically adopts the following technical solutions:

[0010] A method for detecting unmanned aerial vehicle targets based on lightweight feature fusion, the method comprising:

[0011] Step 1: Obtain a training dataset, which includes a public dataset of small targets photographed by drones;

[0012] Step 2: Pre-process the image size of the training dataset and perform mosaic data enhancement. Randomly split and flip any number of images in the training dataset, then flip them and place them at the corresponding segmentation positions. Finally, perform color gamut transformation and affine transformation on the images to obtain training samples.

[0013] Step 3: Based on the training samples, adjust the size of the candidate boxes for network training and calculate the optimal number of candidate boxes required for prediction at each feature extraction layer;

[0014] Step 4: Using the training sample as the input image, the 3×3 convolution module in the CBL module extracts semantic feature information from the input image, and finally increases the channel information of the extracted features to 64 dimensions. The Batch Normalization layer in the CBL module is used to normalize the extracted features; finally, the LeakyRelu activation function in the CBL module is used to enhance the nonlinear factors of the network model.

[0015] Step 5: Use three CSP residual modules CSPBlock1, CSPBlock2, and CSPBlock3 to extract image information features, and output two feature layers of the middle layer CSPBlock2 and CSPBlock3, which are the second feature layer and the third feature layer respectively;

[0016] Step 6: Based on the attention mechanism, the attention weights of the second feature layer and the third feature layer output in step 5 are respectively distributed to update the second feature layer and the third feature layer;

[0017] Step 7: Use the feature fusion module to fuse the receptive field information of different scales on the third feature layer and update the third feature layer;

[0018] Step 8: re-extract features from the extracted feature layer information to generate a first feature layer, and update the second feature layer and the third feature layer;

[0019] Step 9: performing final prediction analysis on the first feature layer, the second feature layer, and the third feature layer;

[0020] Step 10: Debug the network structure hyperparameters from step 4 to step 9, set the network model parameters, and train the network model to obtain the final training model.

[0021] Furthermore, the preprocessing of the image size of the training data set includes:

[0022] The images in the training dataset are resized to 640×640 pixels.

[0023] Furthermore, the step 7, using a feature fusion module to fuse receptive field information of different scales on the third feature layer and update the third feature layer, specifically includes:

[0024] Step 7-1: Use ThreeConv composed of three convolutions to extract the features of the third feature layer obtained in step 6. First, use a 1×1 convolution kernel to reduce the dimension of the third feature layer, then use a 3×3 depth-separable convolution to extract features, and finally use a 1×1 convolution to adjust the output feature layer dimension to update the third feature layer.

[0025] Step 7-2: Based on the cascade feature extraction module SPPF, the SPPF module is composed of three 5×5 maximum pooling layers in cascade, generating receptive field information with the same size as the maximum pooling layers of 5×5, 9×9, and 13×13, respectively. The updated third feature layer is re-extracted by the SPPF module, and the output features on each pooling layer and the input feature information of the SPPF module are concatenated to further update the third feature layer.

[0026] In step 7-3, ThreeConv is used to re-extract features from the updated third feature layer. Finally, the 1×1 convolution in ThreeConv is used to adjust the output feature layer dimension and further update the third feature layer.

[0027] Furthermore, the step 8 of re-extracting features from the extracted feature layer information to generate a first feature layer and update the second feature layer and the third feature layer specifically includes:

[0028] Step 8-1: Add a downsampling feature extraction layer to the network backbone extraction module, and generate the first feature layer by fusing the spatial and position information based on the attention mechanism;

[0029] Step 8-2: performing deep feature fusion on the first feature layer, the second feature layer, and the third feature layer again through a bottom-up path to update the second feature layer and the third feature layer;

[0030] In step 8-3, a top-down path is added to fuse the first feature layer with the second feature layer again, while the second feature layer is not used to fuse with the third feature layer to update the first feature layer.

[0031] Furthermore, step 8-2 further performs deep feature fusion on the first feature layer, the second feature layer, and the third feature layer through a bottom-up path to update the second feature layer and the third feature layer, which specifically includes:

[0032] The third feature layer is upsampled by a factor of 2 using neighboring interpolation to obtain a 40×40 feature map. A 1×1 convolution, normalization, and LeakyRelu activation function are then performed to obtain a second feature layer with 512 channels of the same size. The CA attention mechanism is used on the second feature layer and concat fused with the first feature layer to update the third feature layer.

[0033] The second feature layer is upsampled by a double neighbor interpolation, and then a 1×1 convolution is performed to adjust the number of channels to 258. The CA attention mechanism is used on the second feature layer, and finally it is fused with the first feature layer using Concat to update the second feature layer.

[0034] Furthermore, step 8-3 adds a top-down path to fuse the first feature layer with the second feature layer again, while not fusing the second feature layer with the third feature layer to update the first feature layer, specifically including:

[0035] The first feature layer is subjected to a downsampling convolution with a stride of 2 to obtain a 40×40 feature map, and then a 1×1 convolution, normalization and LeakyRelu activation function are performed to obtain a feature map with the same size and 256 channels. Finally, it is fused with the second feature layer using Concat to update the first feature layer.

[0036] Furthermore, the step 9 of performing final prediction analysis on the first feature layer, the second feature layer, and the third feature layer specifically includes:

[0037] The first, second and third feature layers are respectively subjected to two 1×1 convolutions and a 3×3 convolution in branch one of the Decoupled Head module to extract the classification information of the feature layer. The position information is extracted by using a 3×3 convolution after two 1×1 convolutions in branch two. The confidence information is then extracted by using another 3×3 convolution in branch two, completing the classification and regression tasks of small target detection respectively.

[0038] Furthermore, the step 10, debugging the network structure hyperparameters from the steps 4 to 9, setting the network model parameters, and training the network model to obtain the final training model, specifically includes:

[0039] The number of network training epochs was set to 300 rounds, and transfer learning training was adopted. In the first 60 epochs, the backbone network was frozen, the learning rate was set to 0.001, the pre-trained model was loaded, and the bath size was set to 16. After the next 240 epochs, the backbone network was unfrozen, the learning rate was set to 0.000l, and the bath size was set to 8. After each epoch, the learning rate was reduced to the original 0.937. After 300 epochs of training, the final training model was obtained.

[0040] Furthermore, the method further comprises:

[0041] In step 11, a portion of data is selected from the training data set in step 1 as a test set, and the test set is input into the training model in step 10 to obtain the test results of UAV small target detection.

[0042] Furthermore, images with complex natural scenes, diverse angles, and numerous small objects are selected from the training dataset in step 1 as a test set.

[0043] The present invention has at least the following technical effects:

[0044] (1) The benchmark model of YOLOv4-UAV is the YOLOv4-tiny network. Compared with other one-stage target detection networks, YOLOv4-tiny has the best detection speed while maintaining good detection accuracy, which can well meet the real-time index requirements of UAV target detection.

[0045] (2) After the backbone feature extraction module (backbone) of the YOLOv4-UAV algorithm extracts features, the feature fusion module SPPF is used to re-extract the extracted features at the 32×32 downsampling point, which can effectively increase the receptive field information of the features, solve the multi-scale problem of the target to a certain extent, and reasonably reduce the computational complexity of the model while maintaining the detection accuracy of the model unchanged.

[0046] (3) A new neck feature fusion module CPAN is proposed, which is composed of bottom-up upsampling splicing and top-down downsampling splicing feature layers. By using 8×8 feature detection layer, coordinate attention and depthwise separable convolution, it can effectively solve the problems of low accuracy of small target detection in UAV aerial images, low target detection confidence caused by the difficulty of the detection network to distinguish between background and foreground, and difficulty in real-time detection due to huge network computation.

[0047] (4) The use of the Decoupled Head detection module can effectively solve the detection problem of lack of understanding of spatial information between the target classification task and the positioning structure task, so that the feature information finally extracted has good spatial association, which makes it easier to distinguish between the background and the target to be measured. It can robustly regress the position information of the entire object and has the ability to distinguish between complete targets and partial targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0049] Figure 1 A network block diagram of YOLOv4-UAV and CPAN therein according to an embodiment of the present invention is shown;

[0050] Figure 2 shows a module structure diagram of the SPPF according to an embodiment of the present invention;

[0051] Figure 3 A schematic diagram of the module structure of the CA attention mechanism according to an embodiment of the present invention is shown;

[0052] Figure 4 shows a schematic structural diagram of a Decoupled Head module according to an embodiment of the present invention;

[0053] Figure 5 1 shows a detection result diagram of a DOTA dataset according to an embodiment of the present invention;

[0054] Figure 6 A diagram showing detection results of the Visdrone dataset according to an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0055] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention is described in detail below with reference to the accompanying drawings and specific embodiments. For the steps described herein, if there is no necessary contextual relationship between each other, the order in which they are described as examples herein should not be considered limiting. Those skilled in the art should understand that the order can be adjusted as long as the logic between them is not destroyed, resulting in the entire process being unable to be implemented.

[0056] An embodiment of the present invention provides a drone target detection method based on lightweight feature fusion. The benchmark model YOLOv4-tiny adopted by the method is a one-stage target detection method. The one-stage target detection method can adaptively scale and enhance data on images. It divides the image into multiple grids of equal size in proportion. If the center position of the target is in a certain grid, the grid is responsible for predicting the position and category of the target. It has the advantages of simple model structure and fast detection speed.

[0057] The YOLOv4-UAV network is primarily composed of three modules: the backbone feature extraction module (backbone), the neck feature fusion module (neck), and the detection head module (yolo head). The backbone module utilizes CSPDarnet53-tiny, simplifying the network structure while maintaining the excellent detection accuracy of the CSPDarknet53 module, effectively increasing the model's detection speed. After the YOLOv4-UAV algorithm extracts features using the backbone feature extraction module (backbone), the feature fusion module SPPF is used to re-extract the extracted features at 32×32 downsampling, which can, to a certain extent, address the multi-scale problem of the target. Since the feature fusion module SPPF is equivalent to extracting features using maximum pooling layers of different sizes, the SPPF module can effectively increase the receptive field information of the features. The SPPF module also utilizes receptive fields to design its internal structure. It cascades three 5×5 pooling layers to generate maximum pooling layers of 5×5, 9×9, and 13×13 for feature extraction. This effectively reduces the model's computational complexity while maintaining detection accuracy. Furthermore, to address the challenges of real-time detection in drone aerial images, which often involve numerous small objects and complex backgrounds, a new neck feature fusion module, CPAN, is proposed. This module consists of a bottom-up upsampling and top-down downsampling feature layer concatenation. In addition to using the 32×32 and 16×16 downsampled feature layers from the YOLOv4-tiny network, the CPAN module also uses a feature layer from the 8×8 downsampling network for output. This enhances the CPAN module's ability to extract semantic information from small objects in the image, as lower-level feature layers generally have smaller receptive fields and are therefore more suitable for predicting smaller objects. Furthermore, using Coordinate Attention to construct the CPAN module allows for a weighted fusion of spatial information and channel features, achieving the goal of simultaneously taking into account both channel and position information, enabling the model to better locate objects of interest within the complex background of drone aerial images. Finally, to address the difficulty of real-time detection, depthwise separable convolution (DSC) was used to replace traditional standard convolution, effectively reducing computational parameters. Furthermore, since most drone remote sensing datasets contain small targets, and high-level semantic information has a large receptive field, making it more suitable for detecting large-scale targets, the last 16×16 to 32×32 downsampling layer was cut to limit the increase in parameters of the constructed CPAN module, keeping the network lightweight. Finally, using the Decoupled Head detection module gives the extracted feature information stronger spatial association, thus forming the new YOLOv4-UAV network.

[0058] Specifically, the UAV target detection method based on lightweight feature fusion includes the following steps:

[0059] Step 1: Download the public datasets of small targets photographed by drones, VisDrone and DOTA1.0, and select images with complex natural scenes, diverse angles, and numerous small targets as the test set.

[0060] Step 2: Preprocess the image size of the training dataset, adjust the image size to 640×640 pixels, perform mosaic data augmentation on the dataset in the first 80% of the total training epochs, randomly split and flip four images in the training set, and then flip them. After that, the images are placed in the corresponding segmentation positions, and finally color gamut transformation and affine transformation are performed on the images to obtain the final training samples, effectively increasing the number of training sample sets.

[0061] Step 3: Use the Kmeans algorithm to adjust the size of the candidate boxes (anchors) for network training on the generated new training dataset, and use the Kmeans++ algorithm to calculate the optimal number of candidate boxes required for prediction at each feature extraction layer.

[0062] Figure 1 This is a specific network model diagram of a drone target detection method based on lightweight feature fusion in the present invention. In this embodiment, based on this network model, on the basis of steps 1-3, the following steps are performed:

[0063] In step 4, two CBL modules are used to initially extract image feature information. Semantic feature information is extracted from the input image using the 3×3 convolutional (Conv) modules in the CBL modules, and the channel information of the extracted features is ultimately increased to 64 dimensions. The extracted features are then normalized using the Batch Normalization (BN) layer in the CBL module. Finally, the LeakyRelu activation function in the CBL module is used to enhance the nonlinearity of the network model.

[0064] Step 5, then use Figure 1 The three CSP residual modules CSPBlock1, CSPBlock2, and CSPBlock3 in the image extract the image information features, and output two feature layers feat1 and feat2 of the intermediate layers CSPBlock2 and CSPBlock3 respectively, with the number of channels being 256 and 256 respectively.

[0065] Step 6, use Figure 3The Coordinate Attention (CA) attention mechanism shown in FIG5 assigns attention weights to the two feature layers feat1 and feat2 output in step 5 to generate new feat1 and feat2, and fuses spatial information with channel feature weights, taking into account both channel information and position information.

[0066] Step 7: Use the feature fusion module to fuse the receptive field information of different scales on the feature layer feat2 to obtain a new fused feature layer feat2. The specific implementation is as follows:

[0067] Step 7-1: Use ThreeConv composed of three convolutions to extract the feat2 feature information obtained in step 6. First, use a 1×1 convolution kernel to reduce the dimension of the feature layer, then use a 3×3 depth-separable convolution to extract features, and finally use a 1×1 convolution to adjust the output feature layer dimension to generate a new feature layer feat2.

[0068] In step 7-2, use Figure 2 The cascaded feature extraction module SPPF shown in Figure 1 is composed of three 5×5 max pooling layers in cascade, generating receptive fields of 5×5, 9×9, and 13×13 max pooling layers, respectively. Feature layer feat2 is re-extracted through the SPPF module, and the output features of each pooling layer are concatenated with the input features of the SPPF module to generate a new feature layer feat2 with 1024 channels.

[0069] Step 7-3, use ThreeConv to re-extract features from the feat2 feature layer, and finally use the 1×1 convolution in ThreeConv to adjust the output feature layer dimension to generate a new feature layer feat2;

[0070] Step 8, use Figure 1 The CPAN module in the extractor performs feature re-extraction on the extracted feature layer information. The specific implementation is as follows:

[0071] Step 8-1: Add an 8×8 downsampling feature extraction layer feat0 to the network backbone extraction module, and Figure 3 The Coordinate Attention (CA) attention machine shown in the figure fuses the spatial and position information to generate a new feature layer feat0;

[0072] In step 8-2, the feature layers feat0, feat1, and feat2 obtained in the above steps are further fused through a bottom-up path to obtain feature layers feat0, feat1, and feat2 fused at different scales. The specific implementation is as follows:

[0073] Feat2 is upsampled by a factor of 2 using neighbor interpolation to obtain a 40×40 feature map. Then, a 1×1 convolution, normalization, and LeakyRelu activation function are performed to obtain a feature layer feat2 with a constant size and 512 channels. The CA attention mechanism is then applied to the feature layer feat2, and then feat1 is concatenated. Similarly, feat1 is upsampled by a factor of 2 using neighbor interpolation, and then a 1×1 convolution is performed to adjust the number of channels to 258. The CA attention mechanism is then applied to it, and finally it is concatenated with the feature layer feat0.

[0074] In step 8-3, a top-down path is added to fuse feat0 and feat1 again, without fusing feat1 and feat2, to obtain a new feature layer feat1. The specific implementation is as follows:

[0075] Feat0 is convolved with a downsampling step of 2 to obtain a 40×40 feature map, and then a 1×1 convolution, normalization and LeakyRelu activation function are performed to obtain a feature map with the same size and 256 channels, which is finally fused with feat1 using Concat.

[0076] Step 9, use Figure 4 The Decoupled Head shown performs the final prediction analysis on the final feat0, feat1, and feat2 obtained in the above steps. The specific implementation is as follows:

[0077] The feature layers feat0, feat1, and feat2 are respectively subjected to two 1×1 convolutions and a 3×3 convolution in branch one of the Decoupled Head module to extract the classification information of the feature layer. The position information is extracted by a 3×3 convolution after two 1×1 convolutions in branch two. The confidence information is then extracted by another 3×3 convolution after branch two, completing the classification and regression tasks of small object detection respectively.

[0078] Step 10: Debug the network structure hyperparameters from step 4 to step 9 and set the network model parameters. The number of network training epochs is set to 300 rounds. Transfer learning training is adopted. The backbone network is frozen in the first 60 epochs, the learning rate is set to 0.001, the pre-trained model is loaded, and the bath size is set to 16. After the next 240 epochs, the backbone network is unfrozen and the learning rate is set to 0.000l. The bath size is set to 8, and the learning rate is reduced to the original 0.937 after each epoch. The final training model is obtained after 300 epochs of training.

[0079] In step 11, the test set in step 1 is input into the training model in step 10 to obtain the test results of UAV small target detection.

[0080] The final test results are as follows Figure 5 and Figure 6 As shown in the figure, the detection results of different data sources are respectively shown, which proves that this method can accurately achieve target detection.

[0081] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some or all of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention, and they should all be included in the scope of the claims and description of the present invention.

Claims

1. A UAV target detection method based on lightweight feature fusion, characterized by: The method comprises: Step 1: Obtain the training data set and obtain training samples after preprocessing; Step 2: Using the training sample as the input image, the semantic feature information of the input image is extracted by the 3×3 convolution module in the CBL module, and the channel information of the extracted features is finally increased to 64 dimensions. The extracted features are normalized using the BatchNormalization layer in the CBL module; finally, the LeakyRelu activation function in the CBL module is used to enhance the nonlinear factors of the network model. Step 3: Use three CSP residual modules CSPBlock1, CSPBlock2, and CSPBlock3 to extract image information features, and output two feature layers of the middle layer CSPBlock2 and CSPBlock3, which are the second feature layer and the third feature layer respectively; Step 4: Based on the attention mechanism, the second feature layer and the third feature layer output in step 3 are respectively assigned attention weights to update the second feature layer and the third feature layer; Step 5: Use the feature fusion module to fuse the receptive field information of different scales on the third feature layer and update the third feature layer; Step 6: re-extract features from the extracted feature layer information to generate a first feature layer, and update the second feature layer and the third feature layer; Step 7: performing final prediction analysis on the first feature layer, the second feature layer, and the third feature layer; Step 8: Debug the network structure hyperparameters from step 2 to step 7, set the network model parameters, and train the network model to obtain the final training model; The step 5 specifically includes: Step 5-1: Use ThreeConv composed of three convolutions to extract the features of the third feature layer obtained in step 4. First, use a 1×1 convolution kernel to reduce the dimension of the third feature layer, then use a 3×3 depth-separable convolution to extract features, and finally use a 1×1 convolution to adjust the output feature layer dimension to update the third feature layer. Step 5-2: Based on the cascade feature extraction module SPPF, the SPPF module is composed of three 5×5 maximum pooling layers in cascade, which generate the same receptive field information as the maximum pooling layers of sizes 5×5, 9×9, and 13×13, respectively. The updated third feature layer is re-extracted by the SPPF module, and the output features on each pooling layer and the input feature information of the SPPF module are concatenated to further update the third feature layer. Step 5-3: Use ThreeConv to re-extract features from the updated third feature layer, and finally use the 1×1 convolution in ThreeConv to adjust the output feature layer dimension and further update the third feature layer. The step 6 specifically includes: Step 6-1: Add a downsampling feature extraction layer to the network backbone extraction module, and fuse the spatial and position information based on the attention mechanism to generate the first feature layer; Step 6-2: performing deep feature fusion on the first feature layer, the second feature layer, and the third feature layer again through a bottom-up path to update the second feature layer and the third feature layer; In step 6-3, a top-down path is added to fuse the first feature layer with the second feature layer again, while the second feature layer is not used to fuse with the third feature layer to update the first feature layer.

2. The UAV target detection method based on lightweight feature fusion according to claim 1 is characterized in that: The image size of the training dataset is preprocessed and the image size of the training dataset is adjusted to 640×640 pixels.

3. The UAV target detection method based on lightweight feature fusion according to claim 1 is characterized in that: The step 6-2 specifically includes: The third feature layer is upsampled by a factor of 2 using neighboring interpolation to obtain a 40×40 feature map. A 1×1 convolution, normalization, and LeakyRelu activation function are then performed to obtain a second feature layer with 512 channels of the same size. The CA attention mechanism is used on the second feature layer and concat fused with the first feature layer to update the third feature layer. The second feature layer is upsampled by a double neighbor interpolation, and then a 1×1 convolution is performed to adjust the number of channels to 258. The CA attention mechanism is used on the second feature layer, and finally it is fused with the first feature layer using Concat to update the second feature layer.

4. The UAV target detection method based on lightweight feature fusion according to claim 1 is characterized in that: The step 6-3 specifically includes: The first feature layer is subjected to a downsampling convolution with a stride of 2 to obtain a 40×40 feature map, and then a 1×1 convolution, normalization and LeakyRelu activation function are performed to obtain a feature map with the same size and 256 channels. Finally, it is fused with the second feature layer using Concat to update the first feature layer.

5. The UAV target detection method based on lightweight feature fusion according to claim 1 is characterized in that: The step 7 specifically includes: The first, second and third feature layers are respectively subjected to two 1×1 convolutions and a 3×3 convolution in branch one of the Decoupled Head module to extract the classification information of the feature layer. The position information is extracted by using a 3×3 convolution after two 1×1 convolutions in branch two. The confidence information is then extracted by using another 3×3 convolution in branch two, completing the classification and regression tasks of small target detection respectively.

6. The UAV target detection method based on lightweight feature fusion according to claim 1 is characterized in that: The step 8 specifically includes: The number of network training epochs was set to 300 rounds, and transfer learning training was adopted. In the first 60 epochs, the backbone network was frozen, the learning rate was set to 0.001, the pre-trained model was loaded, and the bath size was set to 16. After the next 240 epochs, the backbone network was unfrozen, the learning rate was set to 0.000l, and the bath size was set to 8. After each epoch, the learning rate was reduced to the original 0.

937. After 300 epochs of training, the final training model was obtained.

7. The UAV target detection method based on lightweight feature fusion according to claim 1 is characterized in that: The method further comprises: In step 9, a portion of the data is selected from the training data set in step 1 as a test set, and the test set is input into the training model in step 8 to obtain the test results of UAV small target detection.

8. The UAV target detection method based on lightweight feature fusion according to claim 7 is characterized in that: From the training dataset in step 1, select images with complex natural scenes, diverse angles, and many small objects as the test set.

Images