Unmanned aerial vehicle small target detection method and device based on iasl-yolo

Abstract

The application discloses a kind of unmanned aerial vehicle small target detection methods based on IASL-YOLO, belong to unmanned aerial vehicle small target detection field based on computer vision.The present application is based on YOLOv8s model, the improved model of the structure that is made of Backbone, Neck and Head;Using labeled training set and validation set to train and optimize the above-mentioned improved model, to obtain optimal improved model;Optimal improved model generated in training stage is pruned using pruning algorithm, generates IASL-YOLO lightweight unmanned aerial vehicle small target detection model;Unlabeled test set is input into IASL-YOLO lightweight unmanned aerial vehicle small target detection model, and the unmanned aerial vehicle small target detection result is output.The model disclosed by the present application not only significantly improves the detection accuracy of unmanned aerial vehicle small target, but also effectively reduces the parameter quantity and calculation amount of the model, realizes the good balance of high precision and low complexity.

Description

Technical Field

[0001] This invention relates to a method and apparatus for small target detection in unmanned aerial vehicles (UAVs) based on IASL-YOLO, belonging to the field of small target detection in UAVs based on computer vision. Background Technology

[0002] With the rapid development of drone technology, drones have been widely used in military, agriculture, logistics, transportation and other fields. However, due to the small target size and high speed of drones, coupled with complex background environments, accurate detection of drones has become a significant challenge in the field of target detection.

[0003] Currently, most mainstream object detection methods are based on deep learning technology. Among them, the YOLO series models (such as YOLOv8 and YOLOv11) are widely used in various object detection tasks due to their fast detection speed and high accuracy. In the task of detecting small targets on UAVs, researchers often improve the above models to enhance the detection accuracy of small targets on UAVs. However, although existing improved models have achieved certain results in terms of detection accuracy, they are often accompanied by a significant increase in the number of model parameters, computational cost, and model size, resulting in a significant increase in model complexity, which is not conducive to deployment on resource-constrained embedded devices.

[0004] Furthermore, while existing lightweight methods reduce the number of model parameters, computational load, and storage requirements to some extent, they typically sacrifice detection accuracy, failing to meet the demands of high-precision detection tasks. Therefore, how to achieve high-precision detection while further reducing the number of model parameters, computational load, and model size remains a key technical challenge in the field of UAV small target detection.

[0005] In view of this, the present invention is hereby proposed. Summary of the Invention

[0006] This invention provides a method and device for small target detection of UAVs based on IASL-YOLO. By integrating the CFE-AFPN module, SIoU loss function and LAMP pruning algorithm, a lightweight UAV small target detection model based on IASL-YOLO is constructed. Based on this model, the accuracy of UAV small target detection is effectively improved and the model is made lighter.

[0007] The technical solution of this invention is:

[0008] According to a first aspect of the present invention, a method for small target detection of unmanned aerial vehicles based on IASL-YOLO is provided, comprising:

[0009] Images of small targets acquired from UAVs were collected, labeled, and then divided into training, validation, and test sets.

[0010] Based on the YOLOv8s model, an improved model is constructed, consisting of a Backbone, Neck, and Head. The Backbone portion of the improved model retains the original YOLOv8s model structure. A CFE-AFPN feature fusion module is introduced to replace the Neck portion of YOLOv8s. The Head portion introduces the SIoU localization loss function, replacing the original CIoU localization loss function. The improved model is trained and optimized using labeled training and validation sets to obtain the optimal improved model.

[0011] A pruning algorithm is used to prune the optimal improved model generated during the training phase to generate the IASL-YOLO lightweight UAV small target detection model; the unlabeled test set is input into the IASL-YOLO lightweight UAV small target detection model, and the UAV small target detection results are output.

[0012] Furthermore, the small target image of the UAV to be detected is input into the IASL-YOLO lightweight UAV small target detection model, which outputs the UAV small target detection result.

[0013] Furthermore, the introduction of the CFE-AFPN feature fusion module to replace the Neck part of YOLOv8s specifically involves: firstly, replacing the Neck part of YOLOv8s with the AFPN module architecture, then replacing the original residual units in AFPN with the C2f-Faster-EMA module to form the CFE-AFPN feature fusion module, and obtaining multi-scale UAV small target feature maps by using the four-level feature maps obtained from the Backbone as input to the CFE-AFPN feature fusion module.

[0014] Furthermore, the C2f-Faster-EMA module specifically involves: dividing the input feature map into two branches after passing it through a convolutional block and then splitting it into two branches; one branch is directly used as part of the subsequent feature fusion, and the other branch is processed by multiple stacked FasterBlock modules, with the output of each FasterBlock module except the last one also being used as part of the subsequent feature fusion to extract more fine-grained features; fusing the two parts of features; and then generating a multi-scale attention feature map after using the EMA attention mechanism and convolutional block operations.

[0015] Furthermore, the pruning algorithm is the LAMP pruning algorithm, specifically comprising: determining the pruning rate; performing layer skipping operations on modules in the original optimal improved model that are not allowed to be pruned; calculating the LAMP score of the weight tensors of each module in the original optimal improved model that are allowed to be pruned; globally pruning the connection with the lowest LAMP score until the required pruning rate is met; and fine-tuning the pruned model using the same labeled training and validation sets as the original optimal improved model to recover the performance lost during the pruning process.

[0016] Furthermore, the modules that are not allowed to be pruned include the head part of the model and the ASFF spatial adaptive network part.

[0017] According to a second aspect of the present invention, an IASL-YOLO-based small target detection device for unmanned aerial vehicles is provided, comprising:

[0018] The first module is used to collect images of small targets from UAVs, and after labeling, divide them into training set, validation set and test set;

[0019] The second module is used to construct an improved model based on the YOLOv8s model, consisting of a Backbone, Neck, and Head. The Backbone part retains the original YOLOv8s model structure. A CFE-AFPN feature fusion module is introduced to replace the Neck part of YOLOv8s. The Head part introduces the SIoU localization loss function to replace its original CIoU localization loss function. The improved model is trained and optimized using labeled training and validation sets to obtain the optimal improved model.

[0020] The third module is used to prune the optimal improved model generated during the training phase using a pruning algorithm to generate the IASL-YOLO lightweight UAV small target detection model; it is used to input the unlabeled test set into the IASL-YOLO lightweight UAV small target detection model and output the UAV small target detection results.

[0021] The beneficial effects of this invention are as follows: This invention constructs an IASL-YOLO lightweight UAV small target detection model by integrating the CFE-AFPN module, SIoU loss function, and LAMP pruning algorithm. By comparing the detection results of this model with existing similar models, the model proposed in this invention not only significantly improves the detection accuracy of UAV small targets, but also effectively reduces the number of model parameters and computational load, achieving a good balance between high accuracy and low complexity. Based on this advantage, it provides effective support for the application of this invention in embedded devices and microcomputers. Attached Figure Description

[0022] Figure 1 This is a flowchart of the present invention;

[0023] Figure 2 This is a structural diagram of the IASL-YOLO lightweight UAV small target detection model proposed in this invention;

[0024] Figure 3 This is a structural diagram of the C2f-Faster-EMA module in the CFE-AFPN feature fusion module of the present invention;

[0025] Figure 4 This is a structural diagram of the FasterBlock in the C2f-Faster-EMA module of the present invention;

[0026] Figure 5 This is a structural diagram of the EMA in the C2f-Faster-EMA module of the present invention;

[0027] Figure 6 This is an image of a small drone target against a complex architectural background.

[0028] Figure 7 For YOLOv8s model pairs Figure 6 The identification result image; UAV: ​​unmanned aerial vehicle; 0.26: the model predicts that the probability of it being an unmanned aerial vehicle is 26%;

[0029] Figure 8 The IASL-YOLO lightweight UAV small target detection model of this invention is for... Figure 6 The identification result image; UAV: ​​unmanned aerial vehicle; 0.64: ​​the model predicts that the probability of it being an unmanned aerial vehicle is 64%;

[0030] Figure 9 This is an image of a small drone target against a complex outdoor background.

[0031] Figure 10 For YOLOv8s model pairs Figure 9 The identification result image; UAV: ​​unmanned aerial vehicle; 0.49: the model predicts that the probability of it being an unmanned aerial vehicle is 49%;

[0032] Figure 11 The IASL-YOLO lightweight UAV small target detection model of this invention is for... Figure 9 The identification result image; UAV: ​​unmanned aerial vehicle; 0.63: the model predicts that the probability of it being an unmanned aerial vehicle is 63%. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be arbitrarily combined with each other.

[0034] Example 1: As Figure 1-11 As shown, according to a first aspect of the present invention, a method for detecting small targets on a UAV based on IASL-YOLO is provided, comprising: acquiring images of small targets on a UAV, and dividing them into training, validation, and test sets after annotation; constructing an improved model based on the YOLOv8s model, the model consisting of a Backbone, a Neck, and a Head; wherein the Backbone part retains the YOLOv8s model structure unchanged; introducing a CFE-AFPN feature fusion module to replace the Neck part of YOLOv8s; introducing a SIoU localization loss function to replace the original CIoU localization loss function in the Head part; training and optimizing the improved model using the labeled training and validation sets to obtain the optimal improved model; pruning the optimal improved model generated during the training phase using a pruning algorithm to generate an IASL-YOLO lightweight UAV small target detection model; inputting an unlabeled test set into the IASL-YOLO lightweight UAV small target detection model and outputting the UAV small target detection result. Further, inputting the image of the small target to be detected on the UAV into the IASL-YOLO lightweight UAV small target detection model and outputting the UAV small target detection result. The structure diagram of the IASL-YOLO lightweight UAV small target detection model mentioned above is as follows: Figure 2 As shown.

[0035] Furthermore, the introduction of the CFE-AFPN feature fusion module to replace the Neck part of YOLOv8s specifically involves: firstly, replacing the Neck part of YOLOv8s with the AFPN module architecture, then replacing the original residual units in AFPN with the C2f-Faster-EMA module to form the CFE-AFPN feature fusion module, and obtaining multi-scale UAV small target feature maps by using the four-level feature maps obtained from the Backbone as input to the CFE-AFPN feature fusion module.

[0036] Specifically, such as Figure 2 As shown, the CFE-AFPN feature fusion module includes a convolutional block, an ASFF spatial adaptive network, and a C2f-Faster-EMA module.

[0037] The ASFF spatial adaptive network specifically involves: obtaining spatial weights for feature maps at multiple levels through 1×1 convolution; multiplying the spatial weights with their corresponding features and then adding them together to obtain a new feature map.

[0038] The C2f-Faster-EMA module specifically works as follows: the input feature map is processed by a convolutional block and then split into two branches: one branch is directly used as part of the subsequent feature fusion, and the other branch is processed by multiple stacked FasterBlock modules, with the output of each FasterBlock module except the last one also being used as part of the subsequent feature fusion to extract more fine-grained features; the two parts of features are fused; then the EMA attention mechanism is used to enhance the feature representation capability through cross-space learning, and a multi-scale attention feature map is generated after further convolutional block operations.

[0039] The C2f-Faster-EMA module introduced in this invention is an improvement on the C2f module. Specifically, Bottleneck in C2f is replaced with FasterBlock, and an EMA attention mechanism is added after fusing the features of the two branches. The structural diagrams of the C2f-Faster-EMA module, the FasterBlock module, and the EMA attention mechanism are shown below. Figure 3 , Figure 4 , Figure 5 As shown.

[0040] Furthermore, the FasterBlock module includes a 3×3 partial convolution and two 1×1 point convolutions. It should be noted that the computational cost of the FasterBlock module is much less than that of the two 3×3 convolutions in the original residual unit of AFPN, and this design further achieves the lightweighting of the model.

[0041] Furthermore, the EMA attention mechanism specifically involves: dividing the input features into three branches: two 1×1 branches and one 3×3 branch; for the two 1×1 branches, a two-dimensional global average pooling operation is used to encode and concatenate the channel information in the two spatial directions; for the 3×3 branch, a 3×3 convolution is used to capture multi-scale feature representations; cross-spatial learning is introduced for the aggregated 1×1 and 3×3 branches; and a dot product operation is used to aggregate the outputs of the two branches to generate a spatial attention map. Specifically, the formula for the two-dimensional global average pooling operation is:

[0042]

[0043] Where (H,W) represents the height and width of the input feature map, and X... c (i,j) represents the number of pixels in the c-th channel of the feature map.

[0044] As can be seen from the above technical solution, the CFE-AFPN feature fusion module of this invention is an improvement on the AFPN module, and further replaces the original residual units in AFPN with the C2f-Faster-EMA module to further enhance feature learning capabilities. Compared with the feature fusion method in YOLOv8s that uses one bottom-up path and one top-down path, the bottom-up path feature fusion method of this invention based on the progressive feature pyramid idea has the advantage of being lightweight. In addition, the CFE-AFPN module outputs four levels of features. Based on the three detection levels (80×80, 40×40, and 20×20) of YOLOv8s, it further introduces a 160×160 detection layer. This improvement makes the model more sensitive to the detection of small targets, thereby effectively improving detection accuracy.

[0045] Furthermore, the SIoU localization loss function consists of angle loss, distance loss, shape loss, and IoU loss. By combining geometric factors (including angle, distance, and shape), it improves the localization accuracy of the target bounding box. The specific calculation formula for the SIoU loss function is as follows:

[0046] Angular loss Λ:

[0047]

[0048] Among them, c h σ represents the height difference between the center points of the ground truth bounding box and the predicted bounding box, and σ is the distance between the center points of the ground truth bounding box and the predicted bounding box.

[0049] Distance loss Δ:

[0050]

[0051] in, The x-coordinates of the center points of the ground truth bounding box and the predicted bounding box are given. c represents the ordinate of the center point of the ground truth bounding box and the predicted bounding box. h1 c is the height of the minimum bounding rectangle between the ground truth bounding box and the predicted bounding box. w ρ is the width of the minimum bounding rectangle between the ground truth bounding box and the predicted bounding box. t Represents ρ x or ρ y .

[0052] Shape loss Ω:

[0053]

[0054] Where (w,h) represents the width and height of the predicted bounding box, (w',h') represents the width and height of the ground truth bounding box, θ represents the attention level of the shape loss, and w t Indicates ww or w h .

[0055] IoU loss:

[0056]

[0057] Where b is the predicted bounding box and b' is the ground truth bounding box.

[0058] SIoU Loss SIoU :

[0059]

[0060] Furthermore, the pruning algorithm is the LAMP pruning algorithm, specifically comprising: determining the pruning rate, i.e., the ratio of the computational cost of the original optimal improved model to the computational cost of the pruned model; performing layer skipping operations on modules in the original optimal improved model that are not allowed to be pruned; calculating the LAMP score of the weight tensors of each module in the original optimal improved model that are allowed to be pruned; globally pruning the connections with the lowest LAMP scores until the required pruning rate is met; and fine-tuning the pruned model using the same labeled training and validation sets as the original optimal improved model to recover the performance lost during the pruning process. The LAMP score of the u-th index of the weight tensor W is calculated as follows:

[0061]

[0062] Where W[u] and W[v] represent the u-th and v-th indices of the weight tensor W, respectively, and v≥u represents all indices from the u-th element to the last element (i.e., u, u+1, u+2, ...). v≥u (W[v]) 2 It is the sum of squares of all weights starting from the u-th element.

[0063] It is important to note that, given the improved model incorporating the CFE-AFPN feature fusion module and the SIoU localization loss function, pruning the model without prior processing would result in the removal of necessary computational components, leading to performance degradation. Therefore, before pruning, a layer skipping operation should be performed. In the model proposed in this invention, the parts requiring layer skipping (i.e., modules that are not allowed to be pruned) are: the model's head section and the ASFF spatial adaptive network section. In the optimal improved model, all modules except those not allowed to be pruned are allowed to be pruned.

[0064] According to a second aspect of the present invention, an IASL-YOLO-based UAV small target detection device is provided, comprising: a first module for acquiring UAV small target images and dividing them into a training set, a validation set, and a test set after annotation; a second module for constructing an improved model based on the YOLOv8s model, the model consisting of a Backbone, a Neck, and a Head; wherein the Backbone part retains the YOLOv8s model structure; a CFE-AFPN feature fusion module is introduced to replace the Neck part of the YOLOv8s model; the Head part introduces an SIoU localization loss function to replace its original CIoU localization loss function; the improved model is trained and optimized using the labeled training and validation sets to obtain the optimal improved model; and a third module for pruning the optimal improved model generated during the training phase using a pruning algorithm to generate an IASL-YOLO lightweight UAV small target detection model; and for inputting an unlabeled test set into the IASL-YOLO lightweight UAV small target detection model and outputting UAV small target detection results.

[0065] According to a third aspect of the present invention, a processor is provided for running a program, wherein the program executes the IASL-YOLO-based UAV small target detection method described above.

[0066] Example 2: The following describes an optional specific implementation process of the present invention based on experimental data. A method for detecting small targets on a UAV based on IASL-YOLO includes the following steps:

[0067] Step 1: The UAV small target detection images used in this invention originate from the publicly available Anti-UAV dataset released by Dalian University of Technology in November 2023. This dataset contains 10,000 UAV target detection images. Based on the original dataset's partitioning method, in this experiment, the images are divided into: training set: 5200 images, validation set: 2600 images, and test set: 2200 images. The training, validation, and test set images are manually labeled in YOLO format using the Labelimg annotation tool.

[0068] Step 2: In this experiment, the hyperparameters for both the model training phase and the fine-tuning training phase after pruning are set as follows: batch size is 24, epoch is 300, image size is 640×640, initial learning rate (lr0) is 0.01, and learning rate factor (lrf) is 0.0001. The software versions used for training are as follows: Python version 3.8.19, PyTorch version 1.12.1, CUDA version 10.2, and YOLOv8 version 8.3.12.

[0069] Step 3: For the models obtained in the model training stage and the fine-tuning training stage after pruning in this experiment, Precision, Recall and mAP50 are used as performance indicators, and Model Size, Parameter and GFLOPs are used as lightweight indicators to evaluate the models.

[0070] The specific explanation is as follows:

[0071] Precision, or accuracy, represents the proportion of samples that the model predicted to be positive, but which were actually positive. The formula for precision is:

[0072]

[0073] TP stands for True Positive Cases, which is the number of samples correctly predicted as drones by the model; FP stands for False Positive Cases, which is the number of samples incorrectly predicted as drones by the model.

[0074] Recall, or the percentage of samples that were actually positive, is the proportion of samples that the model predicts to be positive. The formula for recall is:

[0075]

[0076] FN stands for false negatives, which is the number of samples that the model failed to correctly predict as drones.

[0077] mAP50 (Mean Average Precision at IoU = 0.5) refers to the area under the precision-recall curve of the model when the IoU (Intersection over Union) threshold is 0.5, i.e., the average precision.

[0078] Model Size refers to the size of the model, measured in megabytes (MB). It reflects the resources required for the model to be stored and transmitted.

[0079] A parameter refers to the number of parameters in a model, expressed in units of quantity. It reflects the complexity of the model.

[0080] GFLOPs (Giga Floating-Point Operations per Second) refers to the computational cost of a model, specifically the number of floating-point operations performed per second, measured in billions.

[0081] Step 4: Based on the YOLOv8s detection model, introduce the CFE-AFPN feature fusion module and the SIoU localization loss function to improve the model. Use the training set and validation set from Step 1 to train and optimize the improved model.

[0082] Step 5: Determine the pruning rate (Speed_up). The LAMP pruning algorithm is used to prune the optimal improved model generated during the training phase, generating the IASL-YOLO lightweight UAV small target detection model. It is important to note that the selection of the pruning rate must consider the following two aspects: Firstly, choosing an excessively high pruning rate will result in a significant performance loss and prevent the desired pruning rate from being achieved regardless of further pruning. Secondly, choosing an excessively low pruning rate will result in insufficient model lightweighting and insignificant pruning effects. This invention increases the pruning rate sequentially within a preset range, starting from an initial value and increasing by a preset step size, to obtain models under different pruning rates. For each model under different pruning rates, the optimal pruning rate is determined with the goal of minimizing performance loss while maximizing model pruning. For example, an initial pruning rate of 1 is used, with a step size of 0.5 and a final value of 2.5.

[0083] Step 6: To investigate the optimal pruning rate of the model trained on the selected dataset, this invention designed a comparison of model performance under different pruning rates (Speed_up). The average results of multiple experiments are presented as shown in Table 1.

[0084] Table 1 Comparison of Pruning Rates

[0085]

[0086] The data in Table 1 shows that a pruning rate of 1.0 represents the model without pruning. At a pruning rate of 1.5, compared to the unpruned model, precision improved by 0.8%, recall by 0.5%, and mAP50 by 0.1%. The model size decreased by 61%, the number of parameters decreased by 65%, and the computational cost decreased by 35%. At this pruning rate, the model showed a slight performance improvement while maintaining good lightweight performance. At pruning rates of 2.0 and 2.5, although the model size, number of parameters, and computational cost further decreased, precision, recall, and mAP50 all suffered varying degrees of loss, meaning such models could not maintain good detection performance. Therefore, a pruning rate of 1.5 was chosen for pruning the model.

[0087] To explore the rationality of the IASL-YOLO network, this invention presents the following ablation comparison experiments of the network, as shown in Table 2:

[0088] Table 2 Ablation Experiment

[0089]

[0090] As shown in Table 2, the original YOLOv8s network already demonstrates excellent performance in the field of small target recognition for UAVs. Building upon this, the integration of the CFE-AFPN module (i.e., YOLOv8s+CFE-AFPN) significantly improves model detection performance: Precision increases by 2.0%, Recall by 6.1%, and mAP50 by 3.7%. Simultaneously, in terms of lightweighting, the model size is reduced by 37%, the number of parameters by 39%, and although computational cost increases by 7%, the overall benefits are considerable. Further introduction of the SIoU loss function (i.e., YOLOv8s+CFE-AFPN+SIoU) results in a slight improvement in model performance metrics, specifically a 0.1% increase in Precision, a 0.2% increase in Recall, and a 0.1% increase in mAP50. Finally, the introduction of LAMP pruning technology allows for even greater lightweighting while maintaining the original performance.

[0091] Based on the above implementation example process, in order to verify that the network model of the present invention has an improvement in detection performance compared with the benchmark YOLOv8s model in small target recognition for UAVs, a small target image of a UAV against a complex building background (e.g., ...) is randomly selected from the test set. Figure 6 (as shown) and a small drone target image against a complex outdoor background (such as) Figure 9 As shown in the figure, the recognition results of the above two images using the YOLOv8s model are as follows: Figure 7 , Figure 10 As shown, the recognition results of the IASL-YOLO network model proposed in this invention for the above two images are as follows: Figure 8 , Figure 11 As shown in the figure, in the recognition results of complex architectural background images, both the YOLOv8s model and the IASL-YOLO model predicted drones, with probabilities of 26% and 64%, respectively. In the recognition results of complex outdoor background images, both the YOLOv8s model and the IASL-YOLO model predicted drones, with probabilities of 49% and 63%, respectively. Comparing the recognition results, the IASL-YOLO network model proposed in this invention significantly improves the ability to recognize small drone targets in complex backgrounds compared to the benchmark YOLOv8s model.

[0092] To explore the superiority of the network model of this invention compared to existing methods, this invention reproduced three models currently considered superior in the same field: YOLOv7-tiny, YOLOv10s, and YOLOv11s. These models were trained under the same experimental environment and using the same dataset. The average of multiple experimental results was used to compare the network model of this invention with the network model of this invention. (See Table 3 for details.)

[0093] Table 3 Comparative Experiments

[0094]

[0095] As shown in Table 3, the IASL-YOLO network model of this invention demonstrates superior detection performance and lightweight design in comparative experiments with existing excellent methods YOLOv7-tiny, YOLOv10s, and YOLOv11s. Compared with other existing methods, the IASL-YOLO network model can better fuse features between different levels, and at the same time, it achieves a significant reduction in model size, number of parameters, and computational cost through pruning. This gives it a significant advantage in comparative experiments on small object detection tasks.

[0096] Given the complex characteristics of targets in the UAV small target domain, such as tiny target size, complex image backgrounds, and susceptibility to various interference factors, and considering the high complexity of existing UAV small target detection models, which are unsuitable for some embedded devices and microcomputers, this study proposes a lightweight network model based on improved YOLOv8s—the IASL-YOLO lightweight UAV small target detection model—for UAV small target recognition. Based on the original YOLOv8s, this network first introduces the CFE-AFPN module, which can fuse multi-level features, reducing the semantic gap between non-adjacent level features and effectively improving the detection capability of small targets. Furthermore, the SIoU localization loss function is introduced, combined with geometric factors, to improve the localization accuracy of the target bounding box. For lightweighting, the LAMP pruning algorithm is introduced to prune the model, reducing its size while maintaining high detection performance.

[0097] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A method for small target detection in unmanned aerial vehicles based on IASL-YOLO, characterized in that, include: Images of small targets acquired from UAVs were collected, labeled, and then divided into training, validation, and test sets. Based on the YOLOv8s model, an improved model is constructed, consisting of a Backbone, Neck, and Head. The Backbone portion of the improved model retains the original YOLOv8s model structure. A CFE-AFPN feature fusion module is introduced to replace the Neck portion of YOLOv8s. The Head portion introduces the SIoU localization loss function, replacing the original CIoU localization loss function. The improved model is trained and optimized using labeled training and validation sets to obtain the optimal improved model. The optimal improved model generated during the training phase is pruned using a pruning algorithm to generate the IASL-YOLO lightweight UAV small target detection model; the unlabeled test set is input into the IASL-YOLO lightweight UAV small target detection model, and the UAV small target detection results are output. The image of a small target on a drone to be detected is used as input to the IASL-YOLO lightweight drone small target detection model, and the drone small target detection result is output. The introduction of the CFE-AFPN feature fusion module to replace the Neck part of YOLOv8s specifically involves: firstly, replacing the Neck part of YOLOv8s with the AFPN module architecture; then, replacing the original residual units in AFPN with the C2f-Faster-EMA module to form the CFE-AFPN feature fusion module; and using the four-level feature maps obtained from the Backbone as input to the CFE-AFPN feature fusion module to obtain multi-scale UAV small target feature maps. The CFE-AFPN module outputs four levels of features, which, based on the three detection levels of YOLOv8s, further introduces a 160×160 detection layer. The C2f-Faster-EMA module specifically works as follows: the input feature map is processed through a convolutional block and then split into two branches: one branch is directly used as part of the subsequent feature fusion, and the other branch is processed through multiple stacked FasterBlock modules, with the output of each FasterBlock module except the last one also being used as part of the subsequent feature fusion to extract more fine-grained features; the two parts of features are fused; then, an EMA attention mechanism is used, and after convolutional block operations, a multi-scale attention feature map is generated; The pruning algorithm is the LAMP pruning algorithm, which specifically involves: determining the pruning rate; performing layer skipping operations on modules in the original optimal improved model that are not allowed to be pruned; calculating the LAMP score of the weight tensors of each module in the original optimal improved model that are allowed to be pruned; globally pruning the connection with the lowest LAMP score until the required pruning rate is met; and fine-tuning the pruned model using the same labeled training and validation sets as the original optimal improved model to recover the performance lost during the pruning process. The modules that are not allowed to be pruned include the head part of the model and the ASFF spatial adaptive network part.

2. A UAV small target detection device based on IASL-YOLO for performing the method of claim 1, characterized in that, include: The first module is used to collect images of small targets from UAVs, and after labeling, divide them into training set, validation set and test set; The second module is used to construct an improved model based on the YOLOv8s model, consisting of a Backbone, Neck, and Head. The Backbone part retains the original YOLOv8s model structure. A CFE-AFPN feature fusion module is introduced to replace the Neck part of YOLOv8s. The Head part introduces the SIoU localization loss function to replace its original CIoU localization loss function. The improved model is trained and optimized using labeled training and validation sets to obtain the optimal improved model. The third module is used to prune the optimal improved model generated during the training phase using a pruning algorithm to generate the IASL-YOLO lightweight UAV small target detection model; it is used to input the unlabeled test set into the IASL-YOLO lightweight UAV small target detection model and output the UAV small target detection results.

Images