A method and device for detecting small targets in aerial photography of a UAV based on a FLEX-YOLO network

Abstract

A kind of unmanned aerial vehicle aerial small target detection method and device based on FLEX-YOLO network, its method includes: (1) obtain public unmanned aerial vehicle aerial image dataset and carry out pre-processing;(2) construct FLEX-YOLO network model, the network is based on YOLOv11 network, and MPFP multi-path feedback perception module, JFIN joint feature interaction network, GhostPlus light weight module are designed, and LAMP pruning method is introduced;(3) with FLEX-YOLO network as detection model, it is trained and verified using training set and verification set, generates final detection model;(4) using the detection model of final training completion, with test set as input, FLEX-YOLO model is tested;The present application realizes excellent balance between model performance and efficiency, provides a kind of efficient, reliable solution for unmanned aerial vehicle aerial small target detection task.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to a method and device for detecting small targets in unmanned aerial vehicle (UAV) aerial photography based on a FLEX-YOLO (Feedback-Driven Lightweight and Enhanced Exchange Network) network. Background Art

[0002] In recent years, drones have demonstrated their significant value in target detection thanks to their lightweight, flexible, stealthy, and long flight times. Their ability to flexibly adjust flight altitude and viewing angle enables efficient observation and target tracking across wide areas, significantly improving the efficiency and accuracy of detection tasks and becoming a crucial tool for intelligent visual perception.

[0003] Object detection is a crucial task in computer vision. In recent years, deep learning-based detection algorithms have achieved significant progress in both accuracy and efficiency. Currently, popular object detection methods are categorized into two-stage and single-stage methods. Two-stage algorithms (such as R-CNN and its improved versions) achieve precise localization by generating candidate regions, but this incurs high computational overhead. Single-stage algorithms (such as the YOLO series) achieve faster detection by directly regressing bounding boxes and class probabilities. YOLO models, due to their real-time performance and high efficiency, have become a popular choice for drone aerial photography object detection.

[0004] However, targets in drone aerial photography scenes are typically small in size, densely distributed, frequently occluded, and feature complex backgrounds, which poses a greater challenge to target detection methods. Small targets are easily disturbed by complex backgrounds due to their weak feature expression capabilities, and dense distribution and frequent occlusion further exacerbate the difficulty of detection. Although existing target detection algorithms have made some progress in feature extraction, the detection accuracy of small targets in complex backgrounds remains limited, and the multi-scale feature fusion mechanism also has obvious deficiencies in the interaction of information between non-adjacent layers. In addition, the high computational cost further limits the feasibility of existing methods in practical applications. Therefore, how to improve the performance of small target detection while achieving model lightweighting has become a core issue that needs to be addressed urgently. Summary of the Invention

[0005] The present invention aims to overcome the problems of insufficient feature extraction capability and excessive computational overhead in existing methods for small target detection, and proposes a method and device for detecting small targets in drone aerial photography based on the FLEX-YOLO network.

[0006] By improving the existing network structure and combining efficient feature extraction strategies with lightweight design, the present invention can effectively improve the detection accuracy of small targets while reducing the computational complexity of the model, thereby meeting the requirements of high efficiency and real-time performance in practical applications.

[0007] The present invention achieves the above-mentioned object through the following technical solutions: a method for detecting small targets in drone aerial photography based on FLEX-YOLO network, comprising the following steps:

[0008] S1: Obtain the public drone aerial image dataset Visdrone2019 and perform preprocessing;

[0009] S2: Build a FLEX-YOLO network model. This network uses the YOLOv11 network as the base network. The designed multi-path feedback perception module MPFP is introduced into the backbone network of the YOLOv11. The improved backbone network is obtained by combining the multi-path design with the depthwise separable convolution DWConv (Depthwise Separable Convolution), the bi-level routing attention BRA (Bi-level Routing Attention) and the multi-layer perceptron MLP.

[0010] S3: Improve the neck network of YOLOv11 to a joint feature interaction network (JFIN). By establishing a fully connected information flow channel between different feature layers to enrich the information interaction between non-adjacent layers, an improved neck network is obtained.

[0011] S4: The designed GhostPlus lightweight module is introduced into the backbone network and the neck network of YOLOv11. By replacing the traditional channel splicing operation with a point-by-point addition operation and introducing a "front activation" strategy, the lightweight calculation of the model is achieved;

[0012] S5: Apply the global scoring-based pruning strategy LAMP (Layer-Adaptive Magnitude Pruning) to the constructed model to achieve model compression;

[0013] S6: Use the FLEX-YOLO network as the detection model, train and verify it using the training set and validation set to generate the final detection model;

[0014] S7: Use the final trained detection model and the test set as input to test the FLEX-YOLO model.

[0015] Preferably, in step S1, the pre-processing step includes:

[0016] S11: Convert the original annotation format to YOLO format, where each row of data represents an object, including the object's category number, the center position coordinates x and y of the object in the image, and the width w and height h of the object;

[0017] S12: All images are cropped or padded to a fixed pixel size while maintaining the aspect ratio of the image, and unfilled areas are zero-filled;

[0018] S13: The data set is divided reasonably. The training set is used for model training, the validation set is used for parameter optimization and performance evaluation, and the test set is used for testing the final detection effect.

[0019] Preferably, in step S2, the design method of the multi-path feedback perception module MPFP is:

[0020] S21: DWConv is used instead of standard convolution. The standard convolution is decomposed into a depthwise convolution with a convolution kernel of 3×3. The convolution operation is performed independently on each channel to extract spatial features, and a point-by-point convolution with a convolution kernel of 1×1 is used to linearly combine all channels to fuse the information between channels.

[0021] S22: Introducing the double-layer routing attention BRA, the core of which is to retain the top-k most relevant connections of each region to other regions, and to generate the adjacency matrix A. r Perform the Top-k operation row by row to generate the routing index matrix I r , where each row contains the index of the current region and its k most related regions to avoid redundant calculations;

[0022] I r =TopKIndex(A r ) (1)

[0023] S23: Add a multi-layer perceptron (MLP) to enhance the expressive power of channel features through nonlinear transformation.

[0024] Preferably, in step S3, the design method of the joint feature interaction network JFIN is:

[0025] S31: The feature maps extracted by Backbone in YOLOv11 are marked from large to small in scale as C1 to C5. C1 belongs to the initial stage of feature extraction and does not directly participate in the multi-scale fusion process. First, the mid-level features of C3 and C4 are fused to start the multi-scale fusion sequence:

[0026]

[0027] S32: A fully connected information flow channel is established between different feature layers to enrich the information interaction between non-adjacent layers:

[0028]

[0029] in, represents the output feature after fusion at position i, j in feature layer k, or represents the feature vector extracted from the original feature map, Used to model the feature interactions between C3 and C4, Represents the weight parameter and satisfies the normalization constraint

[0030] S33: Based on JFIN’s multi-scale detection mechanism, a small object detection head with a 160×160 resolution is added to the network structure.

[0031] Preferably, in step S4, the design method of the GhostPlus lightweight module is:

[0032] S41: Use Add operation to replace Concat operation, and replace the traditional channel splicing operation with point-by-point addition operation:

[0033]

[0034] Where s represents the number of feature channels merged by the addition operation;

[0035] S42: The “pre-activation” strategy is introduced, which places the normalization and activation function before the convolution operation, and replaces the traditional ReLU activation function with the SiLU activation function with better nonlinear expression ability.

[0036] Preferably, in step S5, the implementation method of the LAMP pruning algorithm is:

[0037] S51: Based on the weight magnitude, an importance score definition is introduced. For the u-th index in the weight tensor W, the LAMP score is defined as follows:

[0038]

[0039] Where W[u] represents the element with index u in the weight tensor, and the denominator is the sum of the squares of all weights from index u to the end;

[0040] S52: Based on the LAMP score, the relative importance of the target weight in the remaining weights of its layer is measured, and the weights with relatively low LAMP scores are pruned according to the set pruning rate.

[0041] Preferably, in step S6, the FLEX-YOLO network is used as the detection model, and the training set and the validation set are used to train and validate the detection model, and the steps of generating the final detection model include:

[0042] S61: The experimental environment is set up on the Ubuntu 20.04.4 LTS system, based on the PyTorch 2.5.0 framework, and accelerated using CUDA 11.4. The hardware environment is the NVIDIA RTX A6000 GPU with 48G video memory;

[0043] S62: The initial learning rate of the model is set to 0.01, the momentum parameter is set to 0.937, the weight decay coefficient is 0.0005, and the batch size is set to 16;

[0044] S63: Input the preprocessed training set into the FLEX-YOLO network, calculate the classification loss, positioning loss and confidence loss through forward propagation, and then perform backpropagation and update the parameters. The training process continues for 200 epochs.

[0045] S64: After each epoch, the validation set is input into the current model for inference, and the performance indicators are evaluated in real time. After the training is completed, the model weight with the highest mAP on the validation set is selected as the final detection model.

[0046] The second aspect of the present invention relates to a small target detection device for drone aerial photography based on a FLEX-YOLO network, comprising a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the small target detection method for drone aerial photography based on a FLEX-YOLO network of the present invention.

[0047] A third aspect of the present invention relates to a computer-readable storage medium having a program stored thereon. When the program is executed by a processor, the method for detecting small targets in drone aerial photography based on the FLEX-YOLO network of the present invention is implemented.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] Based on the YOLOv11 network, the present invention designs a multi-path feedback perception module MPFP, which significantly enhances the feature extraction capability by fusing the deep separable convolution DWConv, the two-layer routing attention mechanism BRA and the multi-layer perceptron MLP. At the same time, a joint feature interaction network JFIN is designed to achieve more efficient multi-scale feature fusion by establishing a fully connected information flow channel between different feature layers. In addition, the GhostPlus lightweight module is adopted to replace channel splicing by point-by-point addition, and the "pre-activation" strategy is introduced to effectively reduce the computational complexity. In order to further improve the efficiency of the model, the LAMP pruning strategy is used to compress the model, thereby accelerating the inference speed and improving real-time performance. Experimental results show that the present invention surpasses most mainstream target detection algorithms and some advanced small target detection methods on the VisDrone2019 dataset, demonstrating its advantages in performance and efficiency, and providing an accurate, efficient and deployable solution for drone visual perception tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 Schematic diagram of the overall process of the method of the present invention;

[0051] Figure 2 FLEX-YOLO network structure diagram designed for an embodiment of the present invention;

[0052] Figure 3 It is a structural diagram of the MPFP module of the present invention;

[0053] Figure 4 This is the JFIN feature fusion network structure diagram of the present invention;

[0054] Figure 5 It is a structural diagram of the GhostPlus module of the present invention;

[0055] Figure 6 This is a comparison diagram of LAMP pruning of the present invention;

[0056] Figure 7 It is a schematic diagram of the device of the present invention. DETAILED DESCRIPTION

[0057] The present invention is further described below with reference to specific embodiments, but the protection scope of the present invention is not limited thereto:

[0058] Example 1

[0059] like Figure 1 As shown in FIG, a method for detecting small targets in UAV aerial photography based on the FLEX-YOLO network includes the following steps:

[0060] Step S1: Obtain the public drone aerial image dataset Visdrone2019 and perform preprocessing.

[0061] This example uses the VisDrone2019 dataset, collected and compiled by the AISKYEYE team of Tianjin University. This dataset is designed specifically for drone vision tasks and aims to promote the deep integration of drones and visual perception technology. The VisDrone2019 dataset contains 8,629 static images collected from diverse scenes in different cities and rural areas, covering 10 target categories, including pedestrians, vehicles, bicycles, etc., with a total of 343,205 object instances annotated. It presents a variety of visual challenges such as large object size differences, complex backgrounds, and frequent occlusions, making it an important benchmark in the field of drone vision. During data preprocessing, the present invention converts the annotation format of the VisDrone2019 dataset, converting the original annotation format into the YOLO format adapted for the FLEX-YOLO network, where each row of data represents an object, specifically including the target's category number, the target's center position coordinates x and y in the image, and the target's width w and height h. In addition, all images are cropped or padded to a fixed size of 640×640 pixels while maintaining the image's aspect ratio, and unfilled areas are zero-padded. To ensure the effectiveness of model training, the dataset was reasonably divided, with 6,471 images used for the training set, 548 images for the validation set, and 1,610 images for the test set, so as to be used for model training, hyperparameter optimization and performance evaluation, and the final detection effect test.

[0062] Step S2: Construct the FLEX-YOLO network model, the network structure diagram is as follows Figure 2 As shown in the figure. The network is based on the YOLOv11 network. The designed multi-path feedback perception module MPFP is introduced into the backbone network of YOLOv11. Through the multi-path design, the depth-separable convolution DWConv, the double-layer routing attention BRA and the multi-layer perceptron MLP are combined to obtain an improved backbone network. The implementation principle of BRA and the overall architecture of the MPFP module are shown in the figure. Figure 3 shown.

[0063] While the C3K2 module in YOLOv11 has been optimized for feature extraction, it still primarily relies on traditional convolution operations. This local receptive field-based operator has inherent limitations in capturing global information and modeling long-range dependencies, particularly when dealing with small objects and complex scenes. To address this, we designed a multi-path feedback perception module (MPFP). This multi-path design combines depthwise separable convolution (DWConv), dual-layer routing attention (BRA), and multi-layer perceptron (MLP) to comprehensively enhance feature extraction capabilities and achieve efficient object detection in complex scenarios.

[0064] In the MPFP module, DWConv is used to replace the standard convolution, which realizes a significant reduction in computational complexity and parameter quantity by decomposing the standard convolution into a depthwise convolution and a pointwise convolution. Specifically, the depthwise convolution adopts a 3x3 convolution kernel, independently performs convolution operations on each channel, and is used to extract spatial features; the pointwise convolution adopts a 1x1 convolution kernel, linearly combines all channels to fuse the information between channels. This decomposition not only reduces the computational complexity, but also retains the feature extraction ability of the local receptive field.

[0065] At the same time, the MPFP module introduces the BRA dynamic sparse attention mechanism to capture long-distance context information in the spatial dimension. The key of BRA lies in the implementation of its dynamic sparse attention mechanism, which greatly avoids redundant computation through flexible regional routing, and its implementation process is as follows:

[0066] First, for the input feature map X ∈ R H×W×C , it can be divided into SxS regions, each of which contains feature vectors. Then the query, key, and value tensors linearly change as follows:

[0067] Q = X r W q , K = X r W k , V = X r W v (1)

[0068] where X r represents the region-divided feature map, W q , W k , and W v are linear projection matrices used to generate Q, K, and V, respectively.

[0069] On this basis, the association between each region and other regions is determined by constructing a directed graph, and the query and key at the regional level are obtained by performing regional-level average pooling on Q and K, and a adjacency matrix A

[0070] A r = Q r (K r ) T (2)

[0071] BRA performs a row-by-row Top-k operation on A r to generate a routing index matrix Ir , where each row contains the index of the current region and its k most correlated regions for subsequent calculations.

[0072] I r =TopKIndex(A r ) (3)

[0073] Based on the regional routing index matrix I r , BRA implements fine-grained Token-to-Token attention. For the query token of region i, its attention range is limited to I r The key-value pairs in the specified k related areas. Since the related areas are scattered, in order to improve the calculation efficiency, we first use I r Collect keys and values:

[0074] K g =gather(K,I r ),V g =gather(V,I r ) (4)

[0075] in, And use gather to select the key-value pairs from the top-k windows most relevant to it for attention calculation. Then, the attention is calculated by the following formula:

[0076] O=Attention(Q,K g ,V g )+LCE(V) (5)

[0077] Here, LCE(·) indicates parameterization using depthwise convolution.

[0078] In addition, to further optimize the feature expression of the channel dimension, the present invention adds a multi-layer perceptron (MLP) to the MPFP module. The MLP enhances the expressive power of channel features through nonlinear transformation and promotes the interaction of information between channels.

[0079] The MPFP module takes into account both spatial and channel dimension feature expression. By dynamically weighting and fusing global context information with local receptive fields, it not only enhances the model's feature extraction capability for small targets in complex backgrounds, but also effectively alleviates the detection difficulties caused by dense target distribution and frequent occlusion, significantly improving the accuracy and efficiency of feature extraction. Figure 3 shown.

[0080] Step S3: Improve the neck network of YOLOv11 to a joint feature interaction network JFIN. By establishing a fully connected information flow channel between different feature layers to enrich the information interaction between non-adjacent layers, an improved neck network is obtained. Its network structure is as follows: Figure 4 shown.

[0081] The specific fusion method is as follows: The feature maps extracted by Backbone in YOLOv11 are labeled C1 to C5 from largest to smallest scale. C1 belongs to the initial stage of feature extraction and does not directly participate in the multi-scale fusion process. First, mid-level features of C3 and C4 are fused to initiate the multi-scale fusion sequence. This process not only performs weighted integration of features, but also fully shares the high-resolution details of shallow features and the semantic information of deep features through a cross-scale interaction mechanism:

[0082]

[0083] Then, a fully connected information flow channel is established between different feature layers to enrich the information interaction between non-adjacent layers:

[0084]

[0085] in, represents the output feature after fusion at position (i, j) in feature layer k, or represents the feature vector extracted from the original feature map, Used to model the feature interactions between C3 and C4, Represents the weight parameter and satisfies the normalization constraint

[0086] JFIN efficiently connects shallow details with deep semantics, fully integrating multi-scale features and avoiding the weakening of feature information during layer-by-layer transmission. Furthermore, based on JFIN's multi-scale detection mechanism, a new 160×160 resolution small object detection head has been added to the network architecture, further improving its ability to perceive small objects.

[0087] Step S4: The designed GhostPlus lightweight module is introduced into the backbone network and neck network of YOLOv11. The traditional channel splicing operation is replaced by point-by-point addition operation, and the "front activation" strategy is introduced to achieve lightweight calculation of the model.

[0088] YOLOv11 widely uses channel concatenation (concat) operations for feature connection in module design. Although this method can integrate multiple layers of features, it will significantly increase the computational overhead in practical applications, as shown in the following formula, where x is the input feature, s represents the number of feature channels merged by the addition operation, and Φ i (x) represents the feature maps generated by different branches:

[0089] y=Concat([Φ1(x),Φ2(x),...,Φ s-1 (x)]) (9)

[0090] The GhostPlus module designed in this paper replaces the Concat operation with the Add operation. The Add operation replaces the traditional channel splicing operation with a point-by-point addition operation, which greatly reduces redundant calculations and improves the efficiency of feature connection:

[0091]

[0092] At the same time, the module introduces a "pre-activation" strategy, placing the normalization and activation functions before the convolution operation, effectively improving the efficiency of feature extraction and the stability of gradient propagation. In addition, the present invention also replaces the traditional ReLU activation function with the SiLU activation function. Compared with ReLU, SiLU has better nonlinear expression ability and can enhance the feature learning ability of the network without introducing significant additional computational overhead. This "pre-activation" design can effectively avoid the loss of feature information and improve the efficiency of gradient propagation in deep networks. The final optimized module is as follows Figure 5 As shown, the dotted line represents the improved part.

[0093] Step S5: The constructed model is subjected to the global scoring-based pruning strategy LAMP to achieve model compression.

[0094] LAMP is a new scoring metric for global pruning that aims to automatically select layer-by-layer sparsity by minimizing model-level output distortion without requiring additional hyperparameter tuning. LAMP introduces an importance score definition based on weight magnitude. For the u-th index in the weight tensor W, the LAMP score is defined as follows:

[0095]

[0096] Among them, W[u] represents the element with index u in the weight tensor, and the denominator is the sum of the squares of all weights from index u to the end.

[0097] The LAMP score measures the relative importance of the target weight among the remaining weights in its layer, ensuring that larger weights have higher importance scores and are retained first. Through this scoring criterion, LAMP can globally sort the weights of all layers and select the least important weights for pruning. In addition, LAMP has an important property: if the square value of the weight W[u] is greater than the square value of the weight W[v], its corresponding LAMP score will also be higher:

[0098] (W[u]) 2 >(W[v]) 2 →score(u;W)>score(v;W) (12)

[0099] This feature ensures that the importance of weights is highly correlated with their magnitude during LAMP pruning, and automatically balances the layer-by-layer sparsity rate through a global pruning strategy, thereby effectively reducing output distortion. The present invention prunes weights with lower scores based on the LAMP score, thereby achieving model compression and acceleration while minimizing the impact on model performance. This pruning strategy based on global scoring makes the selection of weights more accurate, ensuring a balance between the performance and computational efficiency of the model after pruning. The comparison of the effects before and after pruning is as follows: Figure 6 shown.

[0100] Step S6: Use the FLEX-YOLO network as the detection model, train and verify it using the training set and validation set to generate the final detection model.

[0101] The experimental environment is set up on the Ubuntu 20.04.4 LTS system, based on the PyTorch 2.5.0 framework, and accelerated using CUDA 11.4. The hardware environment is NVIDIA RTX A6000 GPU with 48G video memory.

[0102] The initial learning rate of the model is set to 0.01, the momentum parameter is set to 0.937, the weight decay coefficient is 0.0005, and the batch size is set to 16.

[0103] The VisDrone2019 training set divided in step S1 is input into the FLEX-YOLO network. The model is iteratively trained using the training data until the loss function converges. The training process continues for 200 epochs. After training, the precision (Precision), recall (Recall), and mean average precision (mAP) of the validation set are used as the main evaluation indicators of model performance. The number of model parameters and GFLOPs (Giga Floating Point Operations per Second) are used as the measurement criteria for lightweight evaluation. Based on the verification results, the model hyperparameters are adjusted to finally generate the detection model. The specific definitions of the above performance evaluation indicators are as follows:

[0104]

[0105] In the above formula, true positive (TP) refers to the number of actual positive samples correctly identified as positive, and false positive (FP) refers to the number of negative samples incorrectly classified as positive.

[0106]

[0107] In the above formula, false negative (FN) refers to the number of samples that are actually positive but are incorrectly classified as negative.

[0108]

[0109] In the above formula, AP represents the average precision of a single class, and mAP represents the average precision of all classes.

[0110] Step S7: Using the final trained detection model, taking the test set as input, testing the FLEX-YOLO model.

[0111] Load the weight file trained in step S6 into the detection program, and use the VisDrone2019 test set divided in step S1 to comprehensively test the FLEX-YOLO model. During the testing process, the model performs target detection on each unmanned aerial vehicle aerial image in the test set, outputs detection results containing target categories, bounding boxes and confidence scores, and superimposes the detection results on the original image in a visual form to intuitively show the actual detection effect of the model. At the same time, through the test set, key performance indicators including precision (Precision), recall (Recall), and mean average precision (mAP) are calculated, and the model parameter quantity and GFLOPs are combined to evaluate the lightweight performance of the model. These test results not only comprehensively verify the detection performance and robustness of the FLEX-YOLO model in the unmanned aerial vehicle aerial scene, but also provide important data support for the subsequent optimization and improvement of the model.

[0112] To evaluate the performance of the improved model, the present application conducts a comparative experiment on the model, and the experimental results are shown in Table 1.

[0113] Table 1 Comparison of detection methods on VisDrone2019 validation set

[0114]

[0115] Experimental results show that FLEX-YOLO achieves a mean average performance (mAP) of 45.3% on the VisDrone 2019 validation set, surpassing most mainstream object detection algorithms and achieving a 4.9% improvement over the strongest baseline, YOLOv11-S. With minimal parameters and low GFLOPs, FLEX-YOLO achieves an excellent balance between efficiency and performance. Furthermore, FLEX-YOLO achieves the best overall performance compared to leading small object detection methods.

[0116] Example 2

[0117] like Figure 7 This embodiment relates to a small target detection device for drone aerial photography based on a FLEX-YOLO network, comprising a memory and one or more processors. The memory stores executable code. When the one or more processors execute the executable code, they are used to implement the small target detection method for drone aerial photography based on a FLEX-YOLO network of Example 1.

[0118] Example 3

[0119] This embodiment relates to a computer-readable storage medium having a program stored thereon. When the program is executed by a processor, the method for detecting small targets in drone aerial photography based on the FLEX-YOLO network of Example 1 is implemented.

[0120] The above description is a specific embodiment of the present invention and the technical principles used. If the changes made according to the concept of the present invention do not exceed the spirit covered by the description and drawings, they should still fall within the scope of protection of the present invention.

Claims

1. A method for detecting small targets in UAV aerial photography based on FLEX-YOLO network, characterized in that: The steps include: S1: Obtain the public drone aerial image dataset Visdrone2019 and perform preprocessing; S2: Build a FLEX-YOLO network model. This network uses the YOLOv11 network as the base network. The designed multi-path feedback perception module MPFP is introduced into the backbone network of the YOLOv11. The improved backbone network is obtained by combining the multi-path design with the depthwise separable convolution DWConv, the double-layer routing attention BRA and the multi-layer perceptron MLP. S3: Improve the neck network of YOLOv11 to a joint feature interaction network (JFIN). By establishing a fully connected information flow channel between different feature layers, the improved neck network is obtained to enrich the information interaction between non-adjacent layers. The design method of the Joint Feature Interaction Network JFIN is: S31: The feature maps extracted by Backbone in YOLOv11 are marked from large to small in scale as C1 to C5. C1 belongs to the initial stage of feature extraction and does not directly participate in the multi-scale fusion process. First, the mid-level features of C3 and C4 are fused to start the multi-scale fusion sequence: S32: A fully connected information flow channel is established between different feature layers to enrich the information interaction between non-adjacent layers: in, represents the output feature after fusion at position i, j in feature layer k, or represents the feature vector extracted from the original feature map, Used to model the feature interactions between C3 and C4, Represents the weight parameter and satisfies the normalization constraint S33: Based on JFIN's multi-scale detection mechanism, a small object detection head with a 160×160 resolution is added to the network structure; S4: The designed GhostPlus lightweight module is introduced into the backbone network and the neck network of YOLOv11. By replacing the traditional channel splicing operation with a point-by-point addition operation and introducing a "front activation" strategy, the lightweight calculation of the model is achieved; S5: The constructed model is subjected to the global scoring-based pruning strategy LAMP to achieve model compression; S6: Use the FLEX-YOLO network as the detection model, train and verify it using the training set and validation set to generate the final detection model; S7: Use the final trained detection model and the test set as input to test the FLEX-YOLO model.

2. The method for detecting small targets in drone aerial photography based on FLEX-YOLO network according to claim 1, characterized in that: In step S1, the pre-processing step includes: S11: Convert the original annotation format to YOLO format, where each row of data represents an object, including the object's category number, the center position coordinates x and y of the object in the image, and the width w and height h of the object; S12: All images are cropped or padded to a fixed pixel size while maintaining the aspect ratio of the image, and unfilled areas are zero-filled; S13: The data set is divided reasonably. The training set is used for model training, the validation set is used for parameter optimization and performance evaluation, and the test set is used for testing the final detection effect.

3. The method for detecting small targets in drone aerial photography based on FLEX-YOLO network according to claim 1, characterized in that: In step S2, the design method of the multi-path feedback perception module MPFP is: S21: DWConv is used instead of standard convolution. The standard convolution is decomposed into a depthwise convolution with a convolution kernel of 3×3. The convolution operation is performed independently on each channel to extract spatial features, and a point-by-point convolution with a convolution kernel of 1×1 is used to linearly combine all channels to fuse the information between channels. S22: Introducing the double-layer routing attention BRA, the core of which is to retain the top-k most relevant connections of each region to other regions, and to generate the adjacency matrix A. r Perform the Top-k operation row by row to generate the routing index matrix I r , where each row contains the index of the current region and its k most related regions to avoid redundant calculations; I r =TopKIndex(A r ) (1) S23: Add a multi-layer perceptron (MLP) to enhance the expressive power of channel features through nonlinear transformation.

4. The method for detecting small targets in drone aerial photography based on FLEX-YOLO network according to claim 1, wherein: In step S4, the design method of the GhostPlus lightweight module is: S41: Use Add operation to replace Concat operation, and replace the traditional channel splicing operation with point-by-point addition operation: Where s represents the number of feature channels merged by the addition operation; S42: The "pre-activation" strategy is introduced, which places the normalization and activation function before the convolution operation, and replaces the traditional ReLU activation function with the SiLU activation function with better nonlinear expression ability.

5. The method for detecting small targets in drone aerial photography based on FLEX-YOLO network according to claim 1, characterized in that: In step S5, the implementation method of the LAMP pruning algorithm is: S51: Based on the weight magnitude, an importance score definition is introduced. For the u-th index in the weight tensor W, the LAMP score is defined as follows: Where W[u] represents the element with index u in the weight tensor, and the denominator is the sum of the squares of all weights from index u to the end; S52: Based on the LAMP score, the relative importance of the target weight in the remaining weights of its layer is measured, and the weights with relatively low LAMP scores are pruned according to the set pruning rate.

6. The method for detecting small targets in drone aerial photography based on FLEX-YOLO network according to claim 1, characterized in that: In step S6, the FLEX-YOLO network is used as the detection model, and the training set and the validation set are used to train and validate the detection model. The steps of generating the final detection model include: S61: The experimental environment is set up on the Ubuntu 20.04.4LTS system, based on the PyTorch 2.5.0 framework, and accelerated using CUDA 11.

4. The hardware environment is the NVIDIA RTX A6000 GPU with 48GB of video memory; S62: The initial learning rate of the model is set to 0.01, the momentum parameter is set to 0.937, the weight decay coefficient is 0.0005, and the batch size is set to 16; S63: Input the preprocessed training set into the FLEX-YOLO network, calculate the classification loss, positioning loss and confidence loss through forward propagation, and then perform backpropagation and update the parameters. The training process continues for 200 epochs. S64: After each epoch, the validation set is input into the current model for inference, and the performance indicators are evaluated in real time. After the training is completed, the model weight with the highest mAP on the validation set is selected as the final detection model.

7. A small target detection device for drone aerial photography based on FLEX-YOLO network, characterized in that: The method comprises a memory and one or more processors, wherein the memory stores an executable code, and when the one or more processors execute the executable code, they are used to implement the UAV aerial photography small target detection method based on the FLEX-YOLO network according to any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that A program is stored thereon, and when the program is executed by the processor, the method for detecting small targets in drone aerial photography based on the FLEX-YOLO network described in any one of claims 1 to 6 is implemented.

Images