Unmanned aerial vehicle small target detection method based on YOLOv8 network

Abstract

The invention provides an unmanned aerial vehicle small target detection method based on a YOLOv8 network, and the method comprises the steps: obtaining an image shot by an unmanned aerial vehicle, carrying out the image reconstruction operation based on feature extraction, reflection and iteration of the shot image, obtaining a reconstructed image, and dividing the reconstructed image into a training set and a verification set; constructing an improved YOLOv8 network, and performing target detection on the training set according to the improved YOLOv8 network to obtain a target detection result; marking the verification set, and performing correctness verification on the target detection result and the marked verification set to obtain a verification result; performing optimization training on the improved YOLOv8 network through a verification result to obtain a small target detection network; and detecting the to-be-detected image according to the small target detection network to generate a final detection result. According to the method, the resolution of the unmanned aerial vehicle is improved by reconstructing the image, the YOLOv8 network is optimized and improved, and the small target recognition precision is improved.

Description

Technical Field

[0001] The present invention relates to the field of machine vision technology, and in particular to a method for detecting small targets of unmanned aerial vehicles based on a YOLOv8 network. Background Art

[0002] UAV target detection technology, which combines machine vision and flight control, has rapidly developed into a key technology for automatic identification and tracking of specific targets. Its importance is not only reflected in the military and security fields, but also increasingly prominent in many civilian fields such as agricultural monitoring, environmental protection, traffic monitoring and disaster response. With the advancement of UAV technology, its application has gradually expanded from the military field to commercial and civilian fields, and UAV target detection technology has also emerged. Unlike early UAVs that were mainly used for reconnaissance and surveillance, modern UAVs need to be capable of more complex tasks, such as dynamic target tracking, autonomous navigation and multi-target identification. UAVs are highly portable, flexible and unrestricted by terrain, and can easily capture high-altitude images. However, although these high-altitude images provide a wide field of view and rich information, they also bring challenges such as small target size and terrain occlusion.

[0003] At present, deep learning can be used for target detection and recognition. The target detection and recognition methods used are mainly divided into two categories: one is a two-stage target detection method, such as R-CNN, Fast R-CNN, and Faster R-CNN; the other is a single-stage target detection method, such as SSD and YOLO series; currently, single-stage target detection methods are widely used in target detection and recognition. However, due to the different scales of targets in drone aerial images, many small objects, and high density, small targets have little pixel information and blurred feature textures. Directly applying the cutting-edge version of the YOLO series to drone aerial target detection still has problems such as low detection accuracy and insufficient real-time performance. Therefore, it is very necessary to design a drone small target detection method based on the YOLOv8 network. Summary of the invention

[0004] The purpose of the present invention is to provide a UAV small target detection method based on the YOLOv8 network, so as to improve the resolution of the UAV by reconstructing the image, and optimize and improve the YOLOv8 network to improve the accuracy of small target recognition.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] A method for detecting small targets of unmanned aerial vehicles based on a YOLOv8 network comprises the following steps:

[0007] Acquire images taken by a drone, and perform image reconstruction operations based on feature extraction, reflection, and iteration on the images to obtain reconstructed images, and divide the reconstructed images into a training set and a validation set;

[0008] An improved YOLOv8 network is constructed, and target detection is performed on the training set according to the improved YOLOv8 network to obtain target detection results; the improved YOLOv8 network includes: a backbone network based on wavelet convolution, which is used to extract multi-band features of the training set to obtain multi-scale image features of different dimensions; a neck network based on comprehensive connection, which is used to perform feature fusion operations on multi-scale image features of different dimensions to obtain feature maps;

[0009] The verification set is labeled, and the target detection result is verified with the labeled verification set for correctness to obtain the verification result;

[0010] The improved YOLOv8 network is optimized and trained based on the verification results to obtain a small target detection network;

[0011] The image to be detected is detected according to the small target detection network to generate the final detection result.

[0012] Optionally, an image captured by a drone is obtained, and an image reconstruction operation based on feature extraction, reflection and iteration is performed on the captured image to obtain a reconstructed image, specifically including:

[0013] Constructing a back-projection network, which includes: a feature extraction module, a reflection module and a reconstruction module;

[0014] Performing an upsampling operation on the captured image through a feature extraction module to obtain a first feature;

[0015] Performing a reflection extraction operation on the first feature through a reflection module to obtain a second feature;

[0016] The first feature is fused with the second feature through a reconstruction module to obtain a third feature;

[0017] Performing multiple upsampling operations and reflection extraction operations on the third feature through the feature extraction module and the reflection module to obtain multiple features;

[0018] Multiple features are spliced ​​together through the reconstruction module to obtain a reconstructed image.

[0019] Optionally, the first feature and the second feature are fused through a reconstruction module to obtain a third feature, and the method further includes: calculating the difference between the first feature and the second feature through a huber loss function, and optimizing the reconstruction module through the difference.

[0020] Optionally, an improved YOLOv8 network is constructed, and target detection is performed on the training set according to the improved YOLOv8 network to obtain target detection results, specifically including:

[0021] The traditional convolution in the YOLOv8 backbone network is replaced with a wavelet convolution module, and the multi-scale image features of different dimensions of the training set are obtained through the wavelet convolution module and the pyramid module of the backbone network;

[0022] In the neck network of YOLOv8, a combination of full connection and skip connection is used to perform multi-scale feature fusion operations on multi-scale image features of different dimensions to obtain a feature map;

[0023] Perform regression prediction operations in the head network of YOLOv8 to obtain target detection results.

[0024] Optionally, the traditional convolution in the YOLOv8 backbone network is replaced with a wavelet convolution module, and multi-scale image features of different dimensions of the training set are obtained through the wavelet convolution module and the pyramid module of the backbone network, which also includes:

[0025] The CBAM module is introduced into the backbone network of YOLOv8. The CBAM module includes a spatial attention module and a channel attention module.

[0026] The spatial attention module is used to extract features from the output of the pyramid module to obtain spatial attention features, and the spatial position information of the spatial attention features is obtained. The spatial position information is used as the horizontal coordinate, and the percentage of the spatial attention features in the total output features of the pyramid module is used as the vertical coordinate to construct an attention curve graph, and the spatial attention module is optimized and trained according to the attention curve graph;

[0027] The output of the pyramid module is extracted through the channel attention module to obtain channel features, and each channel feature is assigned a corresponding weight, and the channel attention feature is obtained in a weighted manner;

[0028] The spatial attention features and channel attention features are concatenated to obtain multi-scale image features.

[0029] Optionally, a multi-scale feature fusion operation is performed on multi-scale image features of different dimensions by combining full connection and skip connection in the neck network of YOLOv8 to obtain a feature map, specifically including:

[0030] Fusing the high-level features of the pyramid module with the low-level features of the corresponding pyramid module to obtain a first fused feature;

[0031] The features of the same layer of the pyramid module are fused to obtain the second fused features;

[0032] All the first fusion features and the second fusion features are fused to obtain a feature map.

[0033] Optionally, a regression prediction operation is performed in the head network of YOLOv8 to obtain a target detection result, specifically: a regression prediction operation is performed on the feature map through a Bayesian reasoning method to obtain a regression bounding box and a target label; the regression prediction operation also includes: optimizing the head network through a Focal Loss loss function and an IoU loss function.

[0034] Optionally, the verification set is labeled, and the target detection result is verified against the labeled verification set for correctness to obtain a verification result, specifically including:

[0035] Manually annotate the validation set to obtain true annotations; true annotations include annotation boxes and true labels;

[0036] Compare the regressed bounding boxes and target labels with the true annotations to obtain bounding box confidence and label accuracy.

[0037] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects: the UAV small target detection method based on the YOLOv8 network provided by the present invention comprises: obtaining an image taken by the UAV, and performing an image reconstruction operation based on feature extraction, reflection and iteration on the taken image to obtain a reconstructed image, and dividing the reconstructed image into a training set and a verification set; constructing an improved YOLOv8 network, and performing target detection on the training set according to the improved YOLOv8 network to obtain a target detection result; annotating the verification set, and verifying the correctness of the target detection result with the annotated verification set to obtain a verification result; optimizing and training the improved YOLOv8 network according to the verification result to obtain a small target detection network; detecting the image to be detected according to the small target detection network to generate a final detection result. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. 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 creative labor.

[0039] Figure 1 This is a flow chart of the UAV small target detection method based on the YOLOv8 network of the present invention;

[0040] Figure 2 is a flow chart of the image reconstruction method of the present invention;

[0041] Figure 3The present invention is a flow chart of target detection based on the improved YOLOv8 network;

[0042] Figure 4 It is a flowchart of the improved YOLOv8 neck network of the present invention. DETAILED DESCRIPTION

[0043] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0044] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and understandable, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0045] like Figure 1 As shown, the present invention provides a method for detecting small targets of unmanned aerial vehicles based on a YOLOv8 network, comprising the following steps:

[0046] Step 100: Acquire images taken by a drone, and perform image reconstruction operations based on feature extraction, reflection, and iteration on the taken images to obtain reconstructed images, and divide the reconstructed images into a training set and a validation set.

[0047] Specifically, the specific steps of the image reconstruction operation based on feature extraction, reflection and iteration are as follows: Figure 2 As shown, including:

[0048] Step 101: construct a back projection network, which includes: a feature extraction module, a reflection module and a reconstruction module;

[0049] Step 102: performing an upsampling operation on the captured image through a feature extraction module to obtain a first feature;

[0050] Furthermore, the low-resolution captured image is upsampled by the feature extraction module to map the captured image to high resolution, and the high-resolution image is spatially convolved by the built-in convolution layer to obtain high-resolution features.

[0051] It should be noted that the feature extraction module has built-in ReLU activation function and batch normalization layer. The upsampling operation makes the details of the image more obvious. It can retain the local detail information in the input image while enhancing the image resolution, avoiding blur and distortion, thereby improving the image quality, the convergence speed of the back projection network and the final performance.

[0052] Step 103: performing a reflection extraction operation on the first feature through a reflection module to obtain a second feature;

[0053] Furthermore, the high-resolution features are downsampled by the reflection module to remap the high-resolution features back to low resolution to obtain low-resolution features.

[0054] It should be noted that the downsampling operation is implemented through a large-size filter (such as 8*8 or 12*12), which can effectively reduce the size and amount of image data, thereby reducing the complexity of network calculation and memory usage, so as to improve the training speed of the back-projection network and filter out redundant information, so that the network can focus on key information. Reducing the resolution through downsampling also enhances the back-projection network's ability to learn global features.

[0055] Step 104: The first feature and the second feature are fused by a reconstruction module to obtain a third feature;

[0056] Furthermore, the high-resolution features and low-resolution features are fused by splicing to obtain the fused third feature. The third feature combines information of different scales, contains richer and more accurate content, and enhances the expressiveness of the feature.

[0057] Step 105: performing multiple upsampling operations and reflection extraction operations on the third feature through the feature extraction module and the reflection module to obtain multiple features;

[0058] Furthermore, the third feature is upsampled again, the third feature is mapped back to high resolution, and the second high resolution feature is obtained, and the difference between the second high resolution feature and the previous high resolution feature is downsampled to obtain the second low resolution feature. The difference between the second low resolution feature and the previous low resolution feature is upsampled. The cycle is repeated and stopped after 10 iterations to obtain multiple features. It also includes: calculating the difference between the first feature and the second feature through the huber loss function, and optimizing the reconstruction module through the difference between the two.

[0059] It should be noted that deep learning after multiple iterations can effectively restore image details and improve image quality.

[0060] Step 106: Multiple features are spliced ​​together through a reconstruction module to obtain a reconstructed image.

[0061] Specifically, the reconstructed images were copied into three copies, two of which were used as training sets and one as a validation set.

[0062] It should be noted that in the image reconstruction process, rich details and clear contours are restored from low-resolution images, which significantly improves the resolution and quality of the image. The features extracted during multiple iterations are combined to improve the realism and fineness of the reconstructed image.

[0063] Step 200: construct an improved YOLOv8 network, and perform target detection on the training set according to the improved YOLOv8 network to obtain target detection results; the improved YOLOv8 network includes: a backbone network based on wavelet convolution, the backbone network based on wavelet convolution is used to extract multi-band features of the training set to obtain multi-scale image features of different dimensions; a neck network based on comprehensive connection, the neck network based on comprehensive connection is used to perform feature fusion operation on multi-scale image features of different dimensions to obtain a feature map;

[0064] Specifically, an improved YOLOv8 network is constructed, and target detection is performed on the training set based on the improved YOLOv8 network. The specific steps for obtaining the target detection results are as follows: Figure 3 As shown, including:

[0065] Step 210: replace the traditional convolution in the YOLOv8 backbone network with a wavelet convolution module, and obtain multi-scale image features of different dimensions of the training set through the wavelet convolution module and the pyramid module of the backbone network;

[0066] More specifically, first, the wavelet convolution module uses the two-dimensional Haar wavelet transform to perform multi-level decomposition of the input image to obtain frequency components at different scales. Among them, the Haar wavelet transform uses four filters to decompose the image into four sub-bands: low-frequency components (LL), which are used to capture the low-frequency information of the image, such as the overall shape or outline; horizontal high-frequency components (LH), which are used to capture the horizontal edge information in the image; vertical high-frequency components (HL), which are used to capture the vertical edge information in the image. Diagonal high-frequency components (HH), used to capture the diagonal details of the image. In each level of wavelet transform, the image is downsampled, so that the spatial resolution is halved, but the frequency information is more finely decomposed.

[0067] Then, a small-sized deep convolution kernel is used on each frequency subband to perform convolution operations on each decomposed subband. Since the wavelet transform reduces the spatial resolution of each subband, a small-sized convolution kernel can cover a larger original image area and increase the receptive field. This embodiment uses a small convolution kernel of 3x3 or 5x5 for convolution operations.

[0068] It should be noted that the frequency components at different scales are divided into low-frequency sub-bands (LL) and high-frequency sub-bands (LH, HL and HH). The low-frequency sub-band contains large-scale information of the image, and convolution operations on it help capture global features. The high-frequency sub-band contains local edge and detail information, and convolution operations on it can better capture details and local changes in the image.

[0069] Finally, the inverse wavelet transform re-synthesizes the convolution results of each sub-band into a complete output.

[0070] It should be noted that the inverse wavelet transform is a linear transform, which can reassemble the convolution results losslessly.

[0071] Furthermore, the wavelet convolution module replaces the convolution layers in the original YOLOv8 backbone network layer by layer, including conventional convolution layers and downsampling layers, and then integrates the wavelet convolution features from different layers through the pyramid module, and obtains multi-scale image features of different dimensions by cascading, weighting or splicing multiple layers of features.

[0072] It should be noted that the wavelet convolution module is used to extract features from the input image to obtain multi-level features, which are then effectively combined through the pyramid module, making the network more robust when processing targets of various sizes and improving the accuracy of target detection.

[0073] More specifically, the traditional convolution in the YOLOv8 backbone network is replaced with a wavelet convolution module, and the multi-scale image features of different dimensions of the training set are obtained through the wavelet convolution module and the pyramid module of the backbone network, which also includes:

[0074] The CBAM module is introduced into the backbone network of YOLOv8. The CBAM module includes a spatial attention module and a channel attention module.

[0075] The spatial attention module calculates the average and maximum values ​​of the output features of the pyramid module, generates two spatial feature maps, and then concatenates the two spatial feature maps. After a convolution layer and a Sigmoid activation function, a spatial attention weight map is obtained, and the output of the pyramid module is weighted using the spatial attention weight map to achieve feature extraction and obtain spatial attention features. Align the spatial attention features with the total output features of the pyramid module, obtain the spatial position information of the spatial attention features and the percentage of the spatial attention features in the total output features of the pyramid module, and use the spatial position information as the horizontal coordinate and the percentage of the spatial attention features in the total output features of the pyramid module as the vertical coordinate to construct an attention curve, and optimize the spatial attention module based on the attention curve.

[0076] The channel attention module performs average pooling and maximum pooling operations on the output of the pyramid module to obtain two channel feature vectors. These two feature vectors are concatenated to obtain a new channel feature vector to achieve feature extraction. After a fully connected layer and an activation function, it is mapped to the weight of each channel to obtain a channel weight map. Finally, the channel weight map is used to weight the original pyramid module output to obtain the weighted channel attention feature.

[0077] The spatial attention features and channel attention features are concatenated to obtain multi-scale image features.

[0078] It should be noted that optimizing the model through the attention curve graph enhances the model's learning ability and improves the model's ability to focus on important features. The channel attention module improves the model's attention to important features and enhances the accuracy and robustness of target detection.

[0079] Step 220: In the neck network of YOLOv8, a combination of full connection and skip connection is used to perform multi-scale feature fusion operation on multi-scale image features of different dimensions to obtain a feature map. The specific steps are as follows Figure 4 As shown, including:

[0080] Step 221: Fusing the high-level features of the pyramid module with the low-level features of the corresponding pyramid module to obtain a first fused feature;

[0081] Specifically, the jump connection directly connects the low-level features to a certain layer of the high-level network through operations such as element-level addition or concatenation, which can directly transmit low-level detail information. Then, the low-level features are adjusted to the same spatial dimension as the high-level features through bilinear interpolation or transposed convolution, and then concatenated to obtain the first fused features.

[0082] Step 222: Fusing the features of the same layer of the pyramid module to obtain a second fused feature;

[0083] Specifically, all features obtained through jump connections in the same layer are fused using a weighted average method to obtain a second fused feature, which is then transmitted to a deeper network layer through a long jump connection.

[0084] Step 223: Fuse all the first fusion features and the second fusion features to obtain a feature map.

[0085] Specifically, the last layer of the neck network is the output layer, which is connected to all network layers in the pyramid module in a fully connected manner, and the outputs of all network layers are added and merged to obtain a feature map.

[0086] It should be noted that fusing high-level features with low-level features enables the fused features to carry more semantic relationships and make the feature representation more comprehensive. By integrating features at the same layer, the diversity and expression of features are enhanced, and the model's ability to handle complex scenes is improved. The resulting feature map not only retains rich detail information, but also forms a powerful semantic expression through the combination of higher-level features.

[0087] Step 230: Perform regression prediction operation in the head network of YOLOv8 to obtain the target detection result.

[0088] Specifically, the feature map is regressed and predicted by the Bayesian inference method to obtain the regression bounding box and the target label. The regression prediction operation also includes: optimizing the head network by the Focal Loss loss function and the IoU loss function.

[0089] More specifically, the target label is the classification of the detected target. The calculation formula of the Focal Loss loss function is:

[0090] Focal Loss = -α t (1-p t ) γ log(p t );

[0091] Among them, p t is the classification probability, α t is the balancing factor, and γ is the regulating factor. Based on forward propagation and back propagation, the head network is optimized and trained through the Focal Loss loss function and the IoU loss function.

[0092] It should be noted that the Focal Loss loss function improves the model's attention to complex features, and the IoU loss function enhances the accuracy of regression bounding box prediction.

[0093] This example uses precision, recall, and mAP 50 and mAP 50:95 Four indicators are used to evaluate the performance of the improved YOLOv8 network at different image resolutions (Scale), and the results are shown in Table 1:

[0094]

[0095] All indicators of the improved YOLOv8 network are higher than those of the original YOLOv8 network, which effectively enhances the detection accuracy of the model and improves the small target detection capability of the UAV.

[0096] Step 300: Annotate the verification set, and verify the accuracy of the target detection result with the annotated verification set to obtain the verification result. Specifically, it includes:

[0097] Use the LabelImg tool to manually annotate the validation set and obtain the true annotations, which include the annotation box and the true label.

[0098] Compare the regressed bounding boxes and target labels with the true annotations to obtain bounding box confidence and label accuracy.

[0099] Specifically, the IoU function is used to calculate the ratio of the spatial position of the regression box and the annotation box. If the calculated result is greater than 0.5, the box confidence is marked as 1, and if it does not exceed 0.5, it is marked as 0. The proportion of all results with a box confidence of 1 to all results is calculated to obtain the label accuracy.

[0100] Step 400: Optimize and train the improved YOLOv8 network based on the verification results to obtain a small target detection network;

[0101] Specifically, the position of the frame corresponding to the frame confidence of 0 is corrected according to the annotation frame until the label accuracy reaches a preset accuracy threshold. The preset accuracy threshold in this embodiment is 0.7.

[0102] Step 500: Detect the image to be detected according to the small target detection network to generate a final detection result.

[0103] The beneficial effects of the present invention are as follows:

[0104] 1) The image captured by the drone is reconstructed by image reconstruction operations based on feature extraction, reflection and iteration, which improves the resolution of the image acquired by the drone and makes the features extracted by the network model clearer;

[0105] 2) Combining the wavelet convolution module with the pyramid module makes the network model more robust when dealing with objects of various sizes, thus improving the accuracy of object detection;

[0106] 3) The combination of skip connection and full connection enhances the diversity and expressiveness of the extracted features, and improves the model's ability to handle complex scenes.

[0107] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0108] The present invention uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only used to help understand the method and core ideas of the present invention. At the same time, for those skilled in the art, according to the ideas of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A UAV small target detection method based on YOLOv8 network, characterized in that: The steps include: Acquire images taken by a drone, and perform image reconstruction operations based on feature extraction, reflection, and iteration on the images to obtain reconstructed images, and divide the reconstructed images into a training set and a validation set; An improved YOLOv8 network is constructed, and target detection is performed on the training set according to the improved YOLOv8 network to obtain a target detection result; the improved YOLOv8 network includes: a backbone network based on wavelet convolution, the backbone network based on wavelet convolution is used to perform multi-band feature extraction on the training set to obtain multi-scale image features of different dimensions; a neck network based on comprehensive connection, the neck network based on comprehensive connection is used to perform feature fusion operation on the multi-scale image features of different dimensions to obtain a feature map; Annotating the verification set, and verifying the accuracy of the target detection result and the annotated verification set to obtain a verification result; Optimizing and training the improved YOLOv8 network according to the verification results to obtain a small target detection network; The image to be detected is detected according to the small target detection network to generate a final detection result.

2. The UAV small target detection method based on YOLOv8 network according to claim 1 is characterized in that: Acquire an image taken by a drone, and perform an image reconstruction operation based on feature extraction, reflection and iteration on the image to obtain a reconstructed image, specifically including: Constructing a back-projection network, the back-projection network comprising: a feature extraction module, a reflection module and a reconstruction module; Performing an upsampling operation on the captured image by the feature extraction module to obtain a first feature; Performing a reflection extraction operation on the first feature by the reflection module to obtain a second feature; The first feature and the second feature are fused by the reconstruction module to obtain a third feature; Performing the upsampling operation and the reflection extraction operation on the third feature multiple times by the feature extraction module and the reflection module to obtain multiple features; The multiple features are spliced ​​together by the reconstruction module to obtain the reconstructed image.

3. The UAV small target detection method based on YOLOv8 network according to claim 2 is characterized in that: The first feature and the second feature are fused by the reconstruction module to obtain a third feature, and the method further includes: calculating the difference between the first feature and the second feature by a huber loss function, and optimizing the reconstruction module by the difference.

4. The UAV small target detection method based on YOLOv8 network according to claim 1 is characterized in that: Constructing an improved YOLOv8 network, and performing target detection on the training set according to the improved YOLOv8 network to obtain a target detection result, specifically including: The traditional convolution in the backbone network of YOLOv8 is replaced by a wavelet convolution module, and the multi-scale image features of the different dimensions of the training set are obtained through the wavelet convolution module and the pyramid module of the backbone network; In the neck network of YOLOv8, a combination of full connection and skip connection is used to perform a multi-scale feature fusion operation on the multi-scale image features of different dimensions to obtain a feature map; Perform regression prediction operations in the head network of YOLOv8 to obtain target detection results.

5. The UAV small target detection method based on YOLOv8 network according to claim 4 is characterized in that: The traditional convolution in the backbone network of YOLOv8 is replaced by a wavelet convolution module, and the multi-scale image features of the different dimensions of the training set are obtained through the wavelet convolution module and the pyramid module of the backbone network, further comprising: A CBAM module is introduced into the backbone network of YOLOv8, wherein the CBAM module includes a spatial attention module and a channel attention module; Extracting features from the output of the pyramid module through the spatial attention module to obtain spatial attention features, and obtaining spatial position information of the spatial attention features, taking the spatial position information as the abscissa, and taking the percentage of the spatial attention features in the total output features of the pyramid module as the ordinate, constructing an attention curve graph, and optimizing the spatial attention module according to the attention curve graph; The channel attention module extracts features from the output of the pyramid module to obtain channel features, assigns corresponding weights to each of the channel features, and obtains channel attention features in a weighted manner; The spatial attention features and the channel attention features are concatenated to obtain multi-scale image features.

6. The UAV small target detection method based on YOLOv8 network according to claim 4 is characterized in that: In the neck network of YOLOv8, a combination of full connection and skip connection is used to perform multi-scale feature fusion operation on the multi-scale image features of different dimensions to obtain a feature map, which specifically includes: Fusing the high-level features of the pyramid module with the corresponding low-level features of the pyramid module to obtain a first fused feature; Fusing the features of the same layer of the pyramid module to obtain a second fused feature; All of the first fused features and the second fused features are fused to obtain a feature map.

7. The UAV small target detection method based on YOLOv8 network according to claim 4 is characterized in that: A regression prediction operation is performed in the head network of YOLOv8 to obtain a target detection result, specifically: a regression prediction operation is performed on the feature map through a Bayesian reasoning method to obtain a regression bounding box and a target label; The regression prediction operation also includes: optimizing and training the head network through a FocalLoss loss function and an IoU loss function.

8. The UAV small target detection method based on YOLOv8 network according to claim 7 is characterized in that: The verification set is labeled, and the target detection result is verified with the labeled verification set for correctness to obtain a verification result, which specifically includes: Manually annotating the verification set to obtain true annotations; the true annotations include an annotation box and a true label; The regressed bounding box and target label are compared with the true annotation to obtain the bounding box confidence and label accuracy.

Images