Unmanned aerial vehicle aerial photography small target detection method fusing variable kernel convolution and recursive gating

Abstract

The invention relates to the technical field of computer vision, in particular to an unmanned aerial vehicle aerial small target detection method fusing variable kernel convolution and recursive gating, and is suitable for accurate detection of small targets in unmanned aerial vehicle aerial images. The method includes: acquiring an unmanned aerial vehicle aerial image; preprocessing the image; the optimized YOLOv8 model is utilized to perform small target detection, and the model fuses variable kernel convolution, recursive gating convolution and lightweight convolution structures, adapts to different target size changes, and balances network overhead and performance; and outputting a detection result. By introducing the variable kernel convolution, convolution kernel parameters and sampling shapes are flexible and variable, target diversity in aerial images is effectively dealt with, efficient and accurate small target detection is realized, and powerful support is provided for unmanned aerial vehicle aerial photography technology application.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and particularly to a method for detecting small targets in UAV aerial photography by integrating variable kernel convolution and recursive gating, which is applicable to the accurate detection of small targets in UAV aerial photography images. Background Art

[0002] With the rapid development of UAV technology, UAVs are increasingly widely used in military, agricultural, forestry, disaster area rescue and other fields. Based on the high-quality aerial photography images collected by UAVs, adding intelligent and accurate target detection functions to UAVs has become one of the current research hotspots.

[0003] Target detection is an important research direction in the field of computer vision. With the development of convolutional neural networks, most current target detection methods are dominated by deep learning technology. Among them, the YOLO series of target detection models have received extensive attention due to their high accuracy and high speed. However, in the UAV aerial photography scenario, there are problems such as large differences in target sizes, too small sizes of small targets, complex terrain, and serious occlusion, which seriously affect the accuracy of target detection.

[0004] Although the existing YOLOv8 model can efficiently complete a certain degree of target detection tasks, there are still problems such as missed detection and misdetection in the face of complex scenarios, especially small target detection scenarios. Specifically, the standard convolution kernel in traditional convolutional neural networks has fixed sampling positions and fixed receptive field shapes, with limited feature extraction capabilities for small targets and deformed targets; the feature extraction module in the network structure lacks full utilization of high-order spatial information; at the same time, the increased computational complexity to improve detection accuracy also affects the inference speed of the model.

[0005] In view of the above problems, the present invention proposes a method for detecting small targets in UAV aerial photography by integrating variable kernel convolution and recursive gating. Through three key innovations: variable kernel convolution structure, recursive gating convolution structure, and lightweight convolution structure, the optimization and improvement of the YOLOv8 model are realized, and the detection performance of small targets in UAV aerial photography images is significantly improved. Summary of the Invention

[0006] The present invention aims to solve the problems of incomplete target missed detection and small target detection difficulties in the prior art, and proposes a method for detecting small targets in drone aerial photography that integrates variable kernel convolution and recursive gating. The present invention mainly makes improvements in the following three aspects: In the ACBS structure, variable kernel convolution is used to change the original convolution method, and the convolution kernel is given any number of parameters and any sampling shape, which can adapt to the ever-changing targets while balancing the network overhead and performance; in the C2f_Gnblock structure, recursive convolution is used to construct high-order spatial interactions to obtain high-order feature information, and enhance the feature extraction capability of the original C2f module; in the C2f_GS structure, by combining Conv, DWConv, and shuffle, the network can maintain the accuracy of the original Conv as much as possible while improving the reasoning speed.

[0007] The present invention proposes a method for detecting small targets in UAV aerial photography by integrating variable kernel convolution and recursive gating, comprising:

[0008] Acquisition step, acquiring drone aerial images;

[0009] A processing step of preprocessing the aerial image;

[0010] A detection step, performing small target detection on the preprocessed aerial image using an optimized YOLOv8 model, wherein the optimized YOLOv8 model includes a variable kernel convolution structure, a recursive gated convolution structure, and a lightweight convolution structure;

[0011] Output step: output the target detection results.

[0012] Preferably, the processing steps specifically include:

[0013] Using a labeling tool to label the target in the aerial image to obtain a target detection label;

[0014] The labeled data set is divided into training set, validation set and test set in a ratio of 7:2:1;

[0015] The data set is enhanced to generate training samples with richer features.

[0016] Preferably, the data enhancement includes:

[0017] Using a median filtering method to perform noise reduction processing on the aerial image to improve image quality;

[0018] A series of combinations such as image translation, flipping, random cropping and mosaic enhancement are used to generate a training set with richer features.

[0019] Preferably, the variable kernel convolution structure comprises:

[0020] The convolution kernel offset with the dimension of (B, 2N, H, W) is obtained through convolution operation;

[0021] The corrected sampling coordinates are obtained by adding the offset to the initial sampling coordinates;

[0022] The features at the corresponding positions are obtained through bilinear interpolation and resampling, and the feature representation adapted to the target deformation is obtained.

[0023] Preferably, in the variable kernel convolution structure, the feature extraction of the irregular convolution kernel is performed through:

[0024] Stack the features row - by - row and column - by - column respectively;

[0025] Use column convolution and row convolution to extract the features corresponding to the irregular sampling shape;

[0026] Re - organize the obtained features to form the final feature representation.

[0027] Preferably, the recursive gated convolution structure includes:

[0028] The input features are mapped to initial features through an initial projection;

[0029] Based on the initial features, n - order spatial interaction is realized recursively;

[0030] Feature enhancement is realized by using the gated mechanism of depth - separable convolution and element - wise multiplication;

[0031] The enhanced feature representation is obtained through a final projection.

[0032] Preferably, in the recursive gated convolution structure:

[0033] Each recursive process is expressed as where k = 0, 1,..., n - 1;

[0034] f k is the depth - separable convolution operation, g k is the channel - matching function, and α is the scaling factor for stable training;

[0035] The residual connection method is adopted to alleviate the problems of gradient vanishing and gradient explosion and retain low - level features.

[0036] Preferably, the lightweight convolution structure includes:

[0037] The input is down - sampled through ordinary convolution to obtain high - quality features;

[0038] The result of the ordinary convolution is input into the depth - separable convolution to reduce the computational amount;

[0039] Concatenate the results of ordinary convolution and depthwise separable convolution to obtain mixed features;

[0040] Perform a channel shuffle operation on the mixed features to allow the information generated by standard convolution to penetrate into the information generated by depthwise separable convolution.

[0041] Preferably, in the optimized YOLOv8 model:

[0042] Replace the original CBS module with the variable kernel convolution structure at layers 2, 4, 6, 8, 17, and 20 to adapt to target deformation;

[0043] Replace the original C2f module with the recurrent gated convolution structure at layers 3, 5, and 7 to enhance the feature extraction ability;

[0044] Replace the original C2f module with the lightweight convolution structure at layers 19 and 22 to improve the inference speed.

[0045] Preferably, the detection step further includes:

[0046] Set the initial learning rate to 0.01, the momentum to 0.94, the weight decay to 0.0003, and the training threshold to 0.2;

[0047] Normalize the image size to 640×640;

[0048] Use CIOU as the detection box loss and cross-entropy as the class and object loss;

[0049] Iteratively update the model through backpropagation for 1000 rounds and set the early stopping patience to 20 rounds.

[0050] The beneficial effects of the present invention include: introducing variable kernel convolution to change the ordinary convolution method, endowing the convolution kernel with any number of parameters and any sampling shape, being able to adapt to the size changes of different targets in aerial images, and at the same time balancing network overhead and performance; introducing recurrent gated convolution in the backbone network, using recurrent convolution to construct high-order spatial interactions to obtain high-order feature information, enhancing the feature extraction ability of the original C2f module, and improving the target detection accuracy; lightweight design of the structure in C2f of part of the neck network, combining ordinary convolution, depthwise separable convolution and shuffling operations, ensuring the accuracy of the original ordinary convolution while improving the inference speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 is the overall flow schematic diagram of the method of the present invention;

[0052] Figure 2 is the schematic diagram of the optimized and improved YOLOv8 network architecture of the present invention;

[0053] Figure 3 Schematic diagram of the variable kernel convolution Akconv structure in the present invention;

[0054] Figure 4 Schematic diagram of the C2f_GnBlock structure in the present invention;

[0055] Figure 5 Schematic diagram of the C2f_GSblock structure in the present invention. Detailed implementation manners

[0056] Please refer to the attached Figures 1-5 , and the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0057] As Figure 1 shown, the method for detecting small targets in UAV aerial photography that combines variable kernel convolution and recurrent gating provided by the present invention mainly includes the following steps:

[0058] S1: Acquisition step, acquiring UAV aerial photography images;

[0059] S2: Processing step, preprocessing the aerial photography images;

[0060] S3: Detection step, using the optimized YOLOv8 model to detect small targets in the preprocessed aerial photography images;

[0061] S4: Output step, outputting the target detection results.

[0062] The specific implementation manners of the present invention will be described in detail below in conjunction with the embodiments.

[0063] Regarding the acquisition step

[0064] In a preferred embodiment of the present invention, the acquisition step is specifically to use a UAV to fly at a high altitude to collect aerial photography images of different targets. Preferably, the flight altitude of the UAV can be adjusted according to actual needs, usually between 50 meters and 300 meters, so as to obtain clear and appropriately covered aerial photography images. In the embodiment of the present invention, the UAV is equipped with a high-resolution camera (such as 4K or 8K resolution) to ensure that the acquired images contain sufficient detailed information.

[0065] The processing step specifically includes: using a labeling tool to label the targets in the aerial photography images to obtain target detection labels; dividing the labeled data set into a training set, a validation set, and a test set according to a ratio of 7:2:1; performing data augmentation on the data set to generate training samples with richer features.

[0066] In the specific implementation process, the present invention preferably uses the labelImg tool to label the targets in the collected aerial images. The labeled target categories can be defined according to the actual application scenarios. For example, in a military scenario, it can include tanks, armored vehicles, military trucks, etc.; in an agricultural scenario, it can include crops, agricultural machinery, etc.; in a disaster relief scenario, it can include people, building damages, etc. Through labeling, label information including the category and bounding box coordinates is generated for each target.

[0067] The present invention uses a 7:2:1 ratio to divide the dataset, which is the best ratio verified through multiple experiments. The training set accounts for 70% and can provide sufficient samples for the network to learn; the validation set accounts for 20% for model tuning and preventing overfitting; the test set accounts for 10% for the final performance evaluation. This division ratio can effectively verify the model generalization ability while ensuring the training effect.

[0068] Data augmentation includes: using the median filtering method to perform noise reduction processing on the aerial image to improve the image quality; using a series of combinations such as image translation, flipping, random cropping, and mosaic augmentation to generate a training set with richer features.

[0069] In specific implementation, the present invention first uses the median filtering method to perform noise reduction processing on the aerial image. The mathematical definition of median filtering is as follows:

[0070] y[i,j] = median{x[i+m,j+n]|(m,n)∈W},

[0071] where x[i,j] represents the pixel value of the input image at the position (i,j), y[i,j] represents the pixel value of the output image at the same position, and W represents the filtering window centered on (i,j). The preferred window size of the present invention is 3×3 or 5×5, which can effectively remove noise while retaining image details.

[0072] Then, in the embodiments of the present invention, a variety of data augmentation techniques are adopted, specifically including:

[0073] Image translation: randomly translate the original image in the horizontal and vertical directions, and the translation range is ±15% of the image size;

[0074] Image flipping: including horizontal flipping and vertical flipping, to increase the model's recognition ability for different orientations of the target;

[0075] Random cropping: randomly crop a certain size of area from the original image, and the cropping area ratio is between [0.65, 1.0];

[0076] Mosaic Enhancement: Four different images are stitched together into a new image, and the boundaries are randomly adjusted, significantly increasing the number and diversity of targets in a single image.

[0077] The comprehensive application of these data augmentation methods can generate more diverse training samples, effectively improving the generalization ability of the model and its adaptability to complex scenarios.

[0078] The detection step is the core of the present invention. An optimized YOLOv8 model is used to detect small targets in the preprocessed aerial images. As Figure 2 shown, the present invention has made three aspects of optimization and improvement to the original YOLOv8 model: introducing a variable kernel convolution structure, a recursive gated convolution structure, and a lightweight convolution structure.

[0079] The variable kernel convolution structure includes: obtaining a convolution kernel offset with a dimension of (B, 2N, H, W) through convolution operations; adding the offset to the initial sampling coordinates to obtain the corrected sampling coordinates; and obtaining the features at the corresponding positions through bilinear interpolation and resampling to obtain a feature representation that adapts to the deformation of the target.

[0080] As Figure 3 shown, the variable kernel convolution (Akconv) in the present invention realizes the innovation of the traditional convolution kernel. The traditional convolution kernel has fixed sampling positions and is difficult to adapt to the deformation and scale changes of the target. In contrast, Akconv assigns the initial sampling coordinates to convolutions of any size and adjusts the sample shape with learnable offsets.

[0081] The specific implementation process is as follows: First, obtain the offset of the corresponding kernel through convolution operations, with a dimension of ((B, 2N 2 , H, W)), where B represents the batch size, N represents the side length of the convolution kernel (e.g., N = 3 for a 3×3 convolution kernel), N 2 represents the total number of sampling points (e.g., a 3×3 convolution kernel has 9 sampling points), and each sampling point has two offsets in the x and y directions, so it is 2N 2 , and H and W represent the height and width of the feature map respectively. At the same time, obtain the initial sampling coordinates P0 through the initial coordinate generation algorithm, and then add the offset to the original coordinates (P0 + P n ) to obtain the corrected coordinates. Finally, obtain the features at the corresponding positions through bilinear interpolation and resampling. For a convolution kernel of size K×K, the standard initial coordinates P0 are usually generated by regular grid sampling:

[0082]

[0083] where: i, j ∈ {0, 1,..., K - 1} represent the position indices within the convolution kernel, Denotes the floor value of K divided by 2. This generation method ensures that the sampling points are symmetrically distributed with the center point as the origin. For example, for a 3×3 convolutional kernel, the initial coordinate set is: {(-1,-1),(-1,0),(-1,1),(0,-1),(0,0),(0,1),(1,-1),(1,0),(1,1)}. Different initial coordinate schemes do indeed lead to significant differences in the resampling results. This symmetric distribution centered at the origin is a standard configuration commonly used in variable kernel convolution.

[0084] Feature extraction of the irregular convolutional kernel is achieved by: stacking features row - by - row and column - by - column respectively; using column convolution and row convolution to extract features corresponding to the irregular sampling shape; and reorganizing the obtained features to form the final feature representation.

[0085] Since it is difficult to extract features corresponding to the sampling positions of the irregular convolutional kernel, the method chosen in the present invention is to stack features row - by - row and column - by - column respectively, and use column convolution and row convolution to extract features corresponding to the irregular sampling shape. This method can effectively capture the features of different - shaped targets while maintaining computational efficiency.

[0086] In the network implementation, the present invention replaces the original CBS module in the 2nd, 4th, 6th, 8th, 17th, and 20th layers of YOLOv8 to endow the convolutional kernel with any number of parameters and any sampling shape, which can adapt to the size changes of different targets in aerial images, and at the same time balances the network overhead and performance.

[0087] The recursive gated convolution structure includes: mapping the input features to initial features through an initial projection; based on the initial features, achieving n - order spatial interaction recursively; using a gating mechanism of depth - separable convolution and element - wise multiplication to enhance features; and obtaining the enhanced feature representation through a final projection.

[0088] As Figure 4 shown, the recursive gated convolution (Gnconv) in the present invention is a convolutional structure that can achieve high - order spatial interaction. Ordinary convolution can only capture local spatial information, while distinguishing targets from the background in drone aerial images often requires a wider range of context information.

[0089] In the specific implementation, assume the input is x∈R HW×C , where H and W respectively represent the height and width of the feature map, and C represents the number of channels. The output of the recursive gated convolution can be expressed as:

[0090]

[0091] where φ in is a linear projection operation, corresponding to the 1*1 convolution operation in the figure, used to complete information exchange in the channel dimension.

[0092] Each recursive process is denoted as P k+1 = f k (q k ) ⊙ g k (p k ) / α, where k = 0, 1,..., n - 1; f k is a depthwise separable convolution operation, g k is a channel matching function, and α is a scaling factor for stable training; A residual connection method is adopted to alleviate the problems of gradient vanishing and gradient explosion and retain low-level features.

[0093] In each recursion, f k is a depthwise separable convolution operation, and g k is used to match the number of channels of features during each recursion:

[0094]

[0095] Finally, the output q n of the last recursion is input into the projection layer φ out to obtain the result of G nconv . Although f k (q k ) needs to be calculated once for each k value in the formula, seemingly requiring n independent depthwise separable convolution operations, in actual implementation, through an efficient batch processing technique, all intermediate features can be merged in the channel dimension and then a single extended depthwise separable convolution operation is applied to process all data at once. This optimization technique combines multiple convolution calculations in the recursive process into one batch convolution calculation, significantly reducing the computational overhead and greatly improving the computational efficiency.

[0096] In network design, the structure design of ResNet is mimicked. On the one hand, it alleviates the problems of gradient vanishing and gradient explosion and helps retain low-level features; on the other hand, a deep network is constructed by stacking multiple residual structures, and each residual block can learn the residual between the input and output. This learning method helps the network better capture the feature information in the data.

[0097] In the present invention, the original C2f modules are replaced at the 3rd, 5th, and 7th layers of YOLOv8, and recursive convolution is used to construct high-order spatial interactions to obtain high-order feature information, enhancing the feature extraction ability of the original C2f modules and improving the object detection accuracy.

[0098] The lightweight convolution structure includes: downsampling the input through ordinary convolution to obtain high-quality features; inputting the result of ordinary convolution into depthwise separable convolution to reduce the computational amount; concatenating the results of ordinary convolution and depthwise separable convolution to obtain mixed features; performing a channel shuffle operation on the mixed features to allow the information generated by standard convolution to penetrate into the information generated by depthwise separable convolution.

[0099] As Figure 5 shown, GSConv (Group Shuffle Convolution) first performs downsampling on the input through an ordinary convolution, then inputs the result of the ordinary convolution into DWConv (depthwise separable convolution), concatenates the results of the two convolutions, and finally performs a shuffle operation. The purpose of this is to make the output of depthwise separable convolution as close as possible to that of standard convolution, and use shuffle to allow the information generated by standard convolution to penetrate into each part of the information generated by depthwise separable convolution. This method allows the information from standard convolution to be fully mixed into the output of depthwise separable convolution.

[0100] The calculation method of depthwise separable convolution (DWConv) is as follows:

[0101]

[0102] where X represents the input feature map, K represents the convolution kernel, Y represents the output feature map, i, j represent the spatial positions of the feature map, k represents the channel index, and M, N represent the spatial dimensions of the convolution kernel. Compared with standard convolution, depthwise separable convolution separates spatial convolution and channel convolution, significantly reducing the computational amount.

[0103] The channel shuffle operation can be expressed as:

[0104] Y i,j,(g·s+c)％C = X i,j,g·c+s ,

[0105] where g represents the number of groups, s represents the index within the group, c represents the index between groups, C represents the total number of channels, and % represents the modulo operation. In this way, information between different groups can be effectively exchanged, enhancing the expressive ability of features.

[0106] In the present invention, the original C2f module is replaced at layers 19 and 22 of YOLOv8. These layers correspond to the detection outputs of large and medium-sized targets, and lightweight design is carried out to accelerate the inference speed while maintaining the original accuracy.

[0107] In the optimized YOLOv8 model: at the 2nd, 4th, 6th, 8th, 17th, and 20th layers, the original CBS module is replaced with the variable kernel convolution structure to adapt to target deformation; at the 3rd, 5th, and 7th layers, the original C2f module is replaced with the recursive gated convolution structure to enhance the feature extraction ability; at the 19th and 22nd layers, the original C2f module is replaced with the lightweight convolution structure to improve the inference speed.

[0108] This special interleaved layout innovatively constructs a progressive feature extraction framework of "variable sampling - high - order interaction". In the shallow layers (layers 2 - 8), ACBS and C2f_Gnblock are alternately arranged to form a cascading effect of "adaptive sampling - feature enhancement"; in the deep layers (layers 17 - 22), ACBS and C2f_GS work together to achieve a balance of "accurate features - efficient processing". This layout creates multiple feature flow paths, enhancing the network's expressive ability and training stability.

[0109] The detection steps also include: setting the initial learning rate to 0.01, the momentum to 0.94, the weight decay to 0.0003, and the training threshold to 0.2; normalizing the image size to 640×640; using CIOU as the detection box loss and cross - entropy as the class and object loss; iteratively updating the model through backpropagation for 1000 rounds, and setting the early stopping patience to 20 rounds.

[0110] In the specific implementation, the present invention sets the initial learning rate to 0.01, which is the optimal initial value verified through multiple experiments, capable of avoiding oscillations while ensuring the convergence speed. The momentum is set to 0.94, which helps the model jump out of local optima. The weight decay is set to 0.0003, which can effectively prevent overfitting. The training threshold is set to 0.2, indicating that detection boxes with a confidence level lower than 0.2 will be filtered out. This threshold filters out noise detections while retaining sufficient real targets.

[0111] The present invention normalizes all input images to the size of 640×640. This size balances the requirements of detection accuracy and inference speed. For small target detection, appropriately increasing the input size can provide more pixel information, which is beneficial for small target recognition, but an overly large size will cause a huge increase in the computational amount. Through experimental verification, 640×640 is an ideal balance point.

[0112] The mathematical definition of the CIOU (Complete IoU) loss function is as follows:

[0113]

[0114] Among them, IoU represents the intersection - over - union ratio of the predicted box and the ground - truth box, and ρ(b,b gt) represents the Euclidean distance between the center points of the predicted bounding box and the ground truth bounding box, c represents the diagonal length of the smallest closed region that can simultaneously contain the predicted bounding box and the ground truth bounding box, v represents the aspect ratio consistency metric, and α is a balancing parameter. The CIOU loss considers more geometric factors than the traditional IoU loss, including the overlapping area, the distance between the center points, and the aspect ratio, etc., and can more comprehensively measure the difference between the predicted bounding box and the ground truth bounding box.

[0115] The cross-entropy loss is used for the classification tasks of classes and targets, and its calculation formula is:

[0116]

[0117] where C represents the number of classes, y i represents the ground truth label (0 or 1), and p i represents the predicted probability. The cross-entropy loss can measure the difference between the predicted distribution and the ground truth distribution, and is a commonly used loss function in classification tasks.

[0118] In the present invention, the number of iterations is set to 1000 rounds, which is an empirical value considering the scale of the dataset and the complexity of the model. To prevent overfitting, the early stopping tolerance is set to 20 rounds, that is, if the performance of the validation set does not improve for 20 consecutive rounds, the training process is terminated in advance. This early stopping strategy can effectively avoid the model overfitting the training data.

[0119] The output step mainly outputs the object detection results. In a specific implementation, the detection results include the detected object classes, confidence scores, and bounding box coordinates. The bounding box is usually represented by four values: the x and y coordinates of the upper left corner, and the width and height of the box. Preferably, in the present invention, the detection results are visually displayed on the original image, and the objects of different classes are labeled with bounding boxes of different colors, and the class names and confidence scores are also displayed.

[0120] For the convenience of subsequent applications, the detection results can also be saved in a structured data format, such as JSON or CSV format, containing the detailed information of each detected object. These results can be further used for advanced applications such as object tracking and behavior analysis.

[0121] In a specific application example, the method of the present invention is applied to crop monitoring in an agricultural scenario. The drone flies at an altitude of 150 meters above the farmland to collect high-resolution aerial images. After these images are preprocessed, the optimized YOLOv8 model of the present invention is used for object detection to identify different types of crops, agricultural machinery, and potential pest and disease areas.

[0122] The experimental results show that, compared with the standard YOLOv8 model, the method of the present invention has improved the mean average precision (mAP) of small target detection by 15.8%, reduced the missed detection rate by 22.3%, increased the inference speed by 8.7%, and improved the adaptability to deformed targets by 19.2%. These improvements make agricultural monitoring more accurate and efficient, providing strong technical support for precision agriculture.

[0123] In another application example, the method of the present invention is used in the disaster area rescue scenario. After natural disasters such as earthquakes and floods occur, drones quickly fly over the affected areas to collect aerial images. Through the method of the present invention, it is possible to accurately identify affected people, damaged buildings, and possible rescue channels, providing important reference information for rescue operations. Under complex terrain and adverse weather conditions, the method of the present invention still maintains a high detection accuracy, saving precious time for rescue work.

[0124] In summary, the method for small target detection in UAV aerial photography proposed by the present invention, through the synergistic effect of three key innovations, significantly improves the small target detection performance and demonstrates good applicability and reliability in multiple application scenarios.

[0125] It should be noted that the above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for detecting small targets in UAV aerial photography by integrating variable kernel convolution and recurrent gating, characterized in that, Including: An acquisition step of acquiring aerial images captured by a drone; A processing step of preprocessing the aerial images; A detection step of performing small target detection on the preprocessed aerial images using an optimized YOLOv8 model, where the optimized YOLOv8 model includes a variable kernel convolution structure, a recursive gated convolution structure, and a lightweight convolution structure; An output step of outputting the target detection results.

2. The method according to claim 1, wherein The processing step specifically includes: Using an annotation tool to annotate the targets in the aerial images to obtain target detection labels; Dividing the annotated dataset into a training set, a validation set, and a test set according to a ratio of 7:2:1; Performing data augmentation on the dataset to generate training samples with richer features.

3. The method according to claim 2, wherein The data augmentation includes: Adopting a median filtering method to perform noise reduction processing on the aerial images to improve image quality; Adopting a series of combinations such as image translation, flipping, random cropping, and mosaic augmentation to generate a training set with richer features.

4. The method according to claim 1, characterized in that The variable kernel convolution structure includes: Obtaining a convolution kernel offset with dimensions (B, 2N, H, W) through convolution operations; Adding the offset to the initial sampling coordinates to obtain corrected sampling coordinates; Obtaining features at corresponding positions through bilinear interpolation and resampling to obtain a feature representation adapted to target deformation.

5. The method according to claim 4, wherein In the variable kernel convolution structure, the feature extraction of the irregular convolution kernel is achieved through: Stacking features row by row and column by column respectively; Using column convolution and row convolution to extract features corresponding to irregular sampling shapes; Recombining the obtained features to form the final feature representation.

6. The method according to claim 1, wherein The recursive gated convolution structure includes: Mapping the input features to initial features through an initial projection; Based on the initial features, achieving n-order spatial interaction recursively; Using a gated mechanism of depthwise separable convolution and element-wise multiplication to achieve feature enhancement; Obtaining the enhanced feature representation through a final projection.

7. The method according to claim 6, characterized in that, In the recursive gated convolution structure: Each recursive process is represented as where k = 0, 1, ..., n - 1; f k is the depthwise separable convolution operation, and g k is the channel matching function, and α is the scaling factor for stable training; Adopting a residual connection method to alleviate the problems of gradient vanishing and gradient explosion and retain low-level features.

8. The method according to claim 1, characterized in that, The lightweight convolution structure includes: Performing downsampling on the input through ordinary convolution to obtain high-quality features; Inputting the result of ordinary convolution into depthwise separable convolution to reduce the computational amount; Concatenating the results of ordinary convolution and depthwise separable convolution to obtain mixed features; Performing a channel shuffle operation on the mixed features to enable the information generated by standard convolution to penetrate into the information generated by depthwise separable convolution.

9. The method according to claim 1, characterized in that In the optimized YOLOv8 model: Replacing the original CBS module with the variable kernel convolution structure in the 2nd, 4th, 6th, 8th, 17th, and 20th layers to adapt to target deformation; Replacing the original C2f module with the recursive gated convolution structure in the 3rd, 5th, and 7th layers to enhance the feature extraction ability; Replacing the original C2f module with the lightweight convolution structure in the 19th and 22nd layers to improve the inference speed.

10. The method according to claim 1, characterized in that, The detection step further includes: Setting the initial learning rate to 0.01, the momentum to 0.94, the weight decay to 0.0003, and the training threshold to 0.2; Normalizing the image size to 640×640; Adopting CIOU as the detection box loss and cross entropy as the class and object loss; Iteratively update the model through backpropagation for 1000 iterations and set the early stopping patience to 20 iterations.

Images