Unmanned aerial vehicle small target detection method and device based on IASL-YOLO

Abstract

The invention discloses an unmanned aerial vehicle small target detection method based on IASL-YOLO, and belongs to the field of unmanned aerial vehicle small target detection based on computer vision. The method comprises the following steps of: based on a YOLOv8s model, constructing an improved model of which the structure is composed of Backbone, Neck and Head; training and optimizing the improved model by using a labeled training set and a verification set so as to obtain an optimal improved model; a pruning algorithm is adopted to prune the optimal improved model generated in the training stage, and an IASL-YOLO lightweight unmanned aerial vehicle small target detection model is generated; and inputting the unmarked test set into the IASL-YOLO lightweight unmanned aerial vehicle small target detection model, and outputting an unmanned aerial vehicle small target detection result. The model provided by the invention not only remarkably improves the detection precision of the small target of the unmanned aerial vehicle, but also effectively reduces the parameter quantity and calculation quantity of the model, and realizes good balance between high precision and low complexity.

Description

Technical Field

[0001] The present invention relates to an IASL-YOLO-based unmanned aerial vehicle (UAV) small target detection method and device, and belongs to the field of UAV small target detection based on computer vision. Background Art

[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 size and fast movement speed of drones, as well as the complex background environment, accurate detection of drones has become an important challenge in the field of target detection.

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

[0004] In addition, although the existing lightweight methods reduce the number of model parameters, computational complexity and storage requirements to a certain extent, they usually sacrifice detection accuracy and cannot meet the needs of high-precision detection tasks. Therefore, how to further reduce the number of model parameters, computational complexity and model size while achieving high-precision detection is still a key technical problem in the field of UAV small target detection.

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

[0006] The present invention provides a method and device for detecting small targets of unmanned aerial vehicles based on IASL-YOLO. By integrating the CFE-AFPN module, the SIoU loss function and the LAMP pruning algorithm, an IASL-YOLO lightweight unmanned aerial vehicle small target detection model is constructed. Based on the model, the accuracy of unmanned aerial vehicle small target detection is effectively improved and the model is lightweight.

[0007] The technical solution of the present invention is:

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

[0009] Collect small target images from drones, annotate them and divide them into training set, validation set and test set;

[0010] Based on the YOLOv8s model, an improved model consisting of Backbone, Neck and Head is constructed; the Backbone part of the improved model keeps the YOLOv8s model structure unchanged; the CFE-AFPN feature fusion module is introduced to replace the Neck part of YOLOv8s; the SIoU positioning loss function is introduced to the Head part to replace its original CIoU positioning loss function; the above improved model is trained and tuned using labeled training sets and validation sets to obtain the optimal improved model;

[0011] The pruning algorithm is used to prune the optimal improved model generated in 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 to output the UAV small target detection results.

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

[0013] Furthermore, the CFE-AFPN feature fusion module is introduced to replace the Neck part of YOLOv8s, specifically: first, the Neck part of YOLOv8s is replaced by the AFPN module architecture, and then the C2f-Faster-EMA module is used to replace the original residual unit in AFPN to form a CFE-AFPN feature fusion module, and the 4-level feature maps obtained by Backbone are used as the input of the CFE-AFPN feature fusion module to obtain a multi-scale UAV small target feature map.

[0014] Furthermore, the C2f-Faster-EMA module is specifically as follows: the input feature map is subjected to a convolution block and then a Split operation is performed to divide it into two branches: one branch is directly used as a part of the subsequent feature fusion, and the other branch is processed by multiple stacked FasterBlock modules and the output of each FasterBlock module except the last FasterBlock module is also derived as a part of the subsequent feature fusion to extract more fine-grained features; the above two parts of features are fused; and then the EMA attention mechanism and the convolution block operation are used to generate a multi-scale attention feature map.

[0015] Furthermore, the pruning algorithm is a LAMP pruning algorithm, which specifically includes: determining a pruning rate; performing a layer skipping operation on modules that are not allowed to be pruned in the original optimal improved model; calculating the LAMP score of the weight tensor of each module that is allowed to be pruned in the original optimal improved model; 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 set and validation set as the original optimal improved model to restore the performance lost during the pruning process.

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

[0017] According to a second aspect of the present invention, a UAV small target detection device based on IASL-YOLO is provided, comprising:

[0018] The first module is used to collect small target images of drones, and divide them into training set, verification set and test set after annotation;

[0019] The second module is used to build an improved model consisting of Backbone, Neck and Head based on the YOLOv8s model. The Backbone part keeps the YOLOv8s model structure unchanged. The CFE-AFPN feature fusion module is introduced to replace the Neck part of YOLOv8s. The SIoU positioning loss function is introduced to the Head part to replace its original CIoU positioning loss function. The improved model is trained and tuned using labeled training sets and validation sets to obtain the optimal improved model.

[0020] The third module is used to prune the optimal improved model generated in the training phase using a pruning algorithm to generate an 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 the present invention are as follows: the present invention constructs an IASL-YOLO lightweight UAV small target detection model by integrating the CFE-AFPN module, the SIoU loss function and the LAMP pruning algorithm. By applying the model, the detection effect is compared with that of existing similar models. The model proposed by the present invention not only significantly improves the detection accuracy of small targets of UAVs, but also effectively reduces the number of parameters and the amount of calculation of the model, achieving a good balance between high accuracy and low complexity. Based on this advantage, effective support is provided for the application of the present invention in embedded devices and microcomputers. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 is a flow chart of the present invention;

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

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

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

[0026] Figure 5 It is a structural diagram of 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 building background;

[0028] Figure 7 For the YOLOv8s model Figure 6 Recognition result diagram; UAV: ​​Unmanned Aerial Vehicle; 0.26: The model predicts that the probability of it being a UAV is 26%;

[0029] Figure 8 The IASL-YOLO lightweight UAV small target detection model of the present invention is Figure 6 Identification result diagram; UAV: ​​Unmanned Aerial Vehicle; 0.64: ​​The model predicts that the probability of it being a UAV is 64%;

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

[0031] Fig.10 For the YOLOv8s model Fig. 9 Identification result diagram; UAV: ​​Unmanned Aerial Vehicle; 0.49: The model predicts that the probability of it being a UAV is 49%;

[0032] Fig.11 The IASL-YOLO lightweight UAV small target detection model of the present invention is Fig. 9 Recognition result diagram; UAV: ​​Unmanned Aerial Vehicle; 0.63: The model predicts that the probability of it being a UAV is 63%. DETAILED DESCRIPTION

[0033] In order to make the purpose, technical scheme and advantages of the embodiments of the present invention clearer, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. It should be noted that the embodiments in this application and the features in the embodiments can be combined with each other arbitrarily without conflict.

[0034] Example 1: Figure 1-11 As shown, according to the first aspect of the embodiment of the present invention, a method for detecting small targets of unmanned aerial vehicles based on IASL-YOLO is provided, including: collecting small target images of unmanned aerial vehicles, and dividing them into training sets, verification sets and test sets after annotation; constructing an improved model composed of Backbone, Neck and Head based on the YOLOv8s model; wherein the Backbone part keeps the structure of the YOLOv8s model unchanged; introducing the CFE-AFPN feature fusion module to replace the Neck part of YOLOv8s; introducing the SIoU positioning loss function to replace the original CIoU positioning loss function in the Head part; using the annotated training set and verification set to train and tune the above improved model, so as to obtain the optimal improved model; using the pruning algorithm to prune the optimal improved model generated in the training stage, generating the IASL-YOLO lightweight unmanned aerial vehicle small target detection model; inputting the unlabeled test set into the IASL-YOLO lightweight unmanned aerial vehicle small target detection model, and outputting the detection result of the unmanned aerial vehicle small target. Further, the image of the unmanned aerial vehicle small target to be detected is input into the IASL-YOLO lightweight unmanned aerial vehicle small target detection model, and the detection result of the unmanned aerial vehicle small target is output. The structure diagram of the IASL-YOLO lightweight UAV small target detection model mentioned above is as follows Figure 2 shown.

[0035] Furthermore, the CFE-AFPN feature fusion module is introduced to replace the Neck part of YOLOv8s, specifically: first, the Neck part of YOLOv8s is replaced by the AFPN module architecture, and then the C2f-Faster-EMA module is used to replace the original residual unit in AFPN to form a CFE-AFPN feature fusion module, and the 4-level feature maps obtained by Backbone are used as the input of the CFE-AFPN feature fusion module to obtain a multi-scale UAV small target feature map.

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

[0037] The ASFF spatial adaptive network specifically comprises: performing 1×1 convolution on feature maps of multiple levels to obtain respective spatial weights; multiplying the spatial weights with the respective corresponding features and then adding them to obtain a new feature map.

[0038] The C2f-Faster-EMA module specifically comprises the following steps: the input feature map is subjected to a convolution block and then subjected to a Split operation to be divided into two branches: one branch is directly used as part of the subsequent feature fusion, and the other branch is processed through a plurality of stacked FasterBlock modules and the output of each FasterBlock module except the last FasterBlock module is also derived as part of the subsequent feature fusion to extract more fine-grained features; the above two parts of features are fused; then the EMA attention mechanism is used to enhance the feature representation capability by utilizing cross-space learning, and a multi-scale attention feature map is generated after further convolution block operation.

[0039] The C2f-Faster-EMA module introduced in the present invention is improved based on the C2f module. Specifically, Bottleneck in C2f is replaced by FasterBlock, and the 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 in FIG. Figure 3 , Figure 4 , Figure 5 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 complexity of the FasterBlock module is much smaller than the two 3×3 convolutions in the original residual unit of AFPN. This design further realizes the lightweight of the model.

[0041] Furthermore, the EMA attention mechanism is specifically as follows: the input features are divided 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 connect the channel information in two spatial directions; for the 3×3 branch, a 3×3 convolution is used to capture multi-scale feature expressions; cross-space learning is introduced for the aggregated 1×1 branch and 3×3 branch; the 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] Among them, (H, W) is the height and width of the input feature map, X c (i,j) is the number of pixels in the cth channel of the feature map.

[0044] By applying the above technical solution, it can be known that the CFE-AFPN feature fusion module of the present invention is improved based on the AFPN module, and further adopts the C2f-Faster-EMA module to replace the original residual unit in AFPN to further enhance the feature learning ability. Compared with the feature fusion method of a bottom-up path and a top-down path in YOLOv8s, the feature fusion method of the bottom-up path based on the idea of ​​a progressive feature pyramid of the present invention has the advantage of lightweight. In addition, the CFE-AFPN module outputs four levels of features. On the basis of 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 the detection accuracy.

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

[0046] Angular loss Λ:

[0047]

[0048] Among them, c h is the height difference between the center point of the real box and the predicted box, and σ is the distance between the center point of the real box and the predicted box.

[0049] Distance loss Δ:

[0050]

[0051] in, is the horizontal coordinate of the center point of the real box and the predicted box, is the ordinate of the center point of the real box and the predicted box, c h1 is the minimum bounding rectangle height between the real box and the predicted box, c w is the minimum bounding rectangle width between the real box and the predicted box, ρ t Represents ρ x or y .

[0052] Shape loss Ω:

[0053]

[0054] Among them, (w,h) is the width and height of the predicted box, (w',h') is the width and height of the real box, θ is the degree of attention of the shape loss, w t Indicates ww or h .

[0055] IoU Loss:

[0056]

[0057] Among them, b is the predicted box and b' is the real box.

[0058] SIoU Loss SIoU :

[0059]

[0060] Furthermore, the pruning algorithm is a LAMP pruning algorithm, which specifically includes: determining the pruning rate, that is, the ratio of the computational amount of the original optimal improved model to the computational amount of the pruned model; performing layer skipping operations on the modules that are not allowed to be pruned in the original optimal improved model; calculating the LAMP score of the weight tensor of each module that is allowed to be pruned in the original optimal improved model; globally pruning the connection with the lowest LAMP score until the required pruning rate is met; fine-tuning the pruned model using the same annotated training set and validation set as the original optimal improved model to restore the performance lost during the pruning process. Among them, the LAMP score calculation formula of the u-th index of the weight tensor W is 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 is the sum of the squares of all weights starting from the u-th element.

[0063] It should be noted that, in view of the introduction of the CFE-AFPN feature fusion module and the improved SIoU positioning loss function, if the model is pruned without processing, the necessary parts of the model calculation will be pruned, which will lead to the loss of model performance. Therefore, before the pruning operation, the model should be skipped. In the model proposed by the present invention, the parts that need to skip layers are (i.e., modules that are not allowed to be pruned): the Head part of the model, the ASFF spatial adaptive network part; in the optimal improved model, except for the modules that are not allowed to be pruned, the rest are modules that are allowed to be pruned.

[0064] According to a second aspect of an embodiment of the present invention, a device for detecting small targets of unmanned aerial vehicles based on IASL-YOLO is provided, comprising: a first module, for collecting images of small targets of unmanned aerial vehicles, and dividing them into training sets, verification sets and test sets after annotation; a second module, for constructing an improved model composed of Backbone, Neck and Head based on the YOLOv8s model; wherein the Backbone part keeps the structure of the YOLOv8s model unchanged; a CFE-AFPN feature fusion module is introduced to replace the Neck part of YOLOv8s; a SIoU positioning loss function is introduced into the Head part to replace its original CIoU positioning loss function; the improved model is trained and tuned using annotated training sets and verification sets to obtain an optimal improved model; a third module, for pruning the optimal improved model generated in the training phase using a pruning algorithm to generate an IASL-YOLO lightweight unmanned aerial vehicle small target detection model; and for inputting an unlabeled test set into the IASL-YOLO lightweight unmanned aerial vehicle small target detection model to output a detection result of the unmanned aerial vehicle small target.

[0065] According to a third aspect of an embodiment of the present invention, a processor is provided, wherein the processor is used to run a program, wherein the program executes any one of the above-mentioned methods for detecting small targets of unmanned aerial vehicles based on IASL-YOLO when running.

[0066] Example 2: The following is an optional specific implementation process of the present invention in combination with experimental data. A method for detecting small targets of unmanned aerial vehicles based on IASL-YOLO includes the following steps:

[0067] Step 1. The drone small target detection images used in this invention are derived from the publicly available Anti-UAV dataset released by Dalian University of Technology in November 2023, which covers 10,000 drone target detection images. According to the division method of the original dataset, in this experiment, the above images are divided into: training set: 5,200 images, validation set: 2,600 images, test set: 2,200 images. With the help of Labelimg annotation tool, the training set, validation set, and test set images are manually annotated in YOLO format.

[0068] Step 2. In this experiment, the hyperparameters of the model training stage and the post-pruning fine-tuning training stage of the present invention 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 is 3.8.19, Pytorch version is 1.12.1, CUDA version is 10.2, and YOLOv8 version is 8.3.12.

[0069] Step 3. For the models obtained in the model training stage and the post-pruning fine-tuning training stage 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 instructions are as follows:

[0071] Precision refers to the ratio of samples predicted by the model to those that are actually positive. The formula for precision is:

[0072]

[0073] Among them, TP is the true positive sample, which is the number of samples correctly predicted by the model as drones; FP is the false positive sample, which is the number of samples incorrectly predicted by the model as drones.

[0074] Recall is the recall rate, which refers to the proportion of samples predicted as positive by the model among samples that are actually positive. The formula for the recall rate is:

[0075]

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

[0077] mAP50 (MeanAverage 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, that is, the average precision.

[0078] Model Size refers to the size of the model in megabytes (MB). It reflects the resources required for storage and transmission of the model.

[0079] Parameter refers to the number of parameters in the model, which reflects the complexity of the model.

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

[0081] Step 4. Based on the YOLOv8s detection model, the CFE-AFPN feature fusion module and the SIoU positioning loss function are introduced to improve the model. The improved model is trained and tuned using the training set and validation set in Step 1.

[0082] Step 5, determine the pruning rate (Speed_up), use the LAMP pruning algorithm to prune the optimal improved model generated in the training phase, and generate the IASL-YOLO lightweight UAV small target detection model; it should be noted that for the selection of the pruning rate, the following two issues should be considered: on the one hand, if a pruning rate that is too high is selected, the performance of the model will have a huge performance loss and the set pruning rate will not be achieved no matter how the pruning is performed; on the other hand, if a pruning rate that is too low is selected, the lightweight degree of the model will not be high and the pruning effect will not be obvious. The present invention increases the pruning rate from the initial value to the preset step size within a preset interval to obtain models under different pruning rates. For models under different pruning rates, the optimal pruning rate is determined with the goal of minimizing performance loss while pruning the model to the greatest extent. Exemplarily, 1 is the initial value of the pruning rate, 0.5 is the step size, and 2.5 is the final value.

[0083] Step 6. In order to explore the optimal pruning rate of the model trained by the present invention under the selected data set, the present invention designs a comparative implementation of the model performance under different pruning rates Speed_up, and selects the results of multiple experiments to take the average indicator display, as shown in Table 1:

[0084] Table 1 Pruning rate comparison experiment

[0085]

[0086] According to the data shown in Table 1, when the pruning rate is 1.0, the model is not pruned. When the pruning rate is 1.5, compared with the unpruned model, Precision is improved by 0.8%, Recall is improved by 0.5%, mAP50 is improved by 0.1%, the model size is reduced by 61%, the number of parameters is reduced by 65%, and the amount of calculation is reduced by 35%. At this pruning rate, the model has a slight performance improvement and has good lightweight indicators; when the pruning rate is 2.0 and 2.5, although the model size, number of parameters and amount of calculation are further reduced, Precision, Recall and mAP50 have different degrees of loss, and such a model cannot maintain good detection performance. In summary, a pruning rate of 1.5 is selected to prune the model.

[0087] In order to explore the rationality of the IASL-YOLO network, the present invention provides the following ablation comparison experiment on the network. As shown in Table 2:

[0088] Table 2 Ablation experiment

[0089]

[0090] As shown in Table 2, the original YOLOv8s network has an outstanding performance in the field of small target recognition in drones. On this basis, after integrating the CFE-AFPN module (i.e., YOLOv8s+CFE-AFPN), the model detection performance is significantly improved: Precision increased by 2.0%, Recall increased by 6.1%, and mAP50 increased by 3.7%. At the same time, in terms of lightweight, the model size is reduced by 37%, the number of parameters is reduced by 39%, and although the amount of calculation has increased by 7%, the overall benefit is considerable. Further introducing the SIoU loss function (i.e., YOLOv8s+CFE-AFPN+SIoU), the model performance indicators have been slightly improved, specifically, Precision increased by 0.1%, Recall increased by 0.2%, and mAP50 increased by 0.1%. Further introducing the LAMP pruning technology, the model has achieved a greater degree of lightweighting while maintaining the original performance.

[0091] According to the above implementation example process, in order to verify that the network model of the present invention has improved detection performance in UAV small target recognition compared with the baseline YOLOv8s model, a UAV small target image under a complex building background in the test set is randomly selected (such as Figure 6 ) and a small target image of a drone under a complex outdoor background (as shown in Fig. 9 As shown in Figure 2, the recognition results of the above two images using the YOLOv8s model are as follows: Figure 7 , Fig.10 As shown in FIG. 1 , the recognition results of the IASL-YOLO network model proposed in the present invention for the above two images are as follows Figure 8 , Fig.11 As shown. Among them, in the recognition results of complex building background images, both the YOLOv8s model and the IASL-YOLO model predicted drones, and the probabilities of the two models predicting that they are drones are 26% and 64% respectively; in the recognition results of complex outdoor background images, both the YOLOv8s model and the IASL-YOLO model predicted drones, and the probabilities of the two models predicting that they are drones are 49% and 63% respectively. Comparing the recognition results, the IASL-YOLO network model proposed in the present invention has a significantly improved recognition ability of small drone targets in complex backgrounds compared with the baseline YOLOv8s model.

[0092] In order to explore the superiority of the network model of the present invention compared with the existing methods, the present invention reproduced the three models of YOLOv7-tiny, YOLOv10s and YOLOv11s, which are currently superior in the same field, and trained them with the same data set under the same experimental environment. The average value of the indicators of multiple experimental results was selected to conduct a comparative experiment with the network model of the present invention. As shown in Table 3:

[0093] Table 3 Comparative experiment

[0094]

[0095] As shown in Table 3, the IASL-YOLO network model of the present invention shows superior detection performance and lightness in the comparative experiment with the existing excellent methods YOLOv7-tiny, YOLOv10s and YOLOv11s. Compared with other existing methods, the IASL-YOLO network model can better fuse the features between the layers, and at the same time, it can greatly reduce the model size, parameter amount and calculation amount through pruning, which makes it have a significant advantage in the comparative experiment of small target detection tasks.

[0096] In view of the complex characteristics of targets in the field of small targets of drones, such as small target size, complex image background and possible influence of multiple interference factors, as well as the high complexity of the model in the existing small target detection methods of drones, which cannot be applied to some embedded devices and microcomputers, this study proposes a lightweight network model based on the improvement of YOLOv8s-IASL-YOLO lightweight drone small target detection model for the recognition of small targets of drones. Based on the original YOLOv8s, the network first introduces the CFE-AFPN module, which can fuse multi-level features, reduce the semantic gap between non-adjacent level features, and effectively improve the detection ability of small targets. On top of this, the SIoU positioning loss function is further introduced, combined with geometric factors, to improve the positioning accuracy of the target bounding box. In terms of lightweight, the LAMP pruning algorithm is introduced to prune the model, which reduces the size of the model while ensuring the high detection performance of the model.

[0097] The specific implementation modes of the present invention are described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above implementation modes, and various changes can be made within the knowledge scope of ordinary technicians in this field without departing from the purpose of the present invention.

Claims

1. A UAV small target detection method based on IASL-YOLO, characterized in that: include: Collect small target images from drones, annotate them and divide them into training set, validation set and test set; Based on the YOLOv8s model, an improved model consisting of Backbone, Neck and Head is constructed; the Backbone part of the improved model keeps the YOLOv8s model structure unchanged; the CFE-AFPN feature fusion module is introduced to replace the Neck part of YOLOv8s; the SIoU positioning loss function is introduced to the Head part to replace its original CIoU positioning loss function; the above improved model is trained and tuned using labeled training sets and validation sets to obtain the optimal improved model; The pruning algorithm is used to prune the optimal improved model generated in 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 to output the UAV small target detection results.

2. The UAV small target detection method based on IASL-YOLO according to claim 1, characterized in that: It is used to input the image of the small drone target to be detected into the IASL-YOLO lightweight drone small target detection model and output the drone small target detection result.

3. The CFE-AFPN feature fusion module is introduced to replace the Neck part of YOLOv8s. Specifically, the Neck part of YOLOv8s is replaced by the AFPN module architecture. Then, the C2f-Faster-EMA module is used to replace the original residual unit in AFPN to form a CFE-AFPN feature fusion module. The 4-level feature maps obtained by Backbone are used as the input of the CFE-AFPN feature fusion module to obtain a multi-scale UAV small target feature map.

4. The UAV small target detection method based on IASL-YOLO according to claim 3 is characterized in that: The C2f-Faster-EMA module specifically comprises the following steps: the input feature map is passed through a convolution block and then split into two branches by a Split operation: one branch is directly used as part of the subsequent feature fusion, and the other branch is processed by multiple stacked FasterBlock modules, and the output of each FasterBlock module except the last FasterBlock module is also derived as part of the subsequent feature fusion to extract more fine-grained features; the above two parts of features are fused; and then the EMA attention mechanism and the convolution block operation are used to generate a multi-scale attention feature map.

5. The UAV small target detection method based on IASL-YOLO according to claim 1, characterized in that: The pruning algorithm is a LAMP pruning algorithm, which specifically includes: determining a pruning rate; performing a layer skipping operation on the modules that are not allowed to be pruned in the original optimal improved model; and calculating the LAMP score of the weight tensor of each module that is allowed to be pruned in the original optimal improved model; The connections with the lowest LAMP scores are pruned globally until the required pruning rate is met; the pruned model is fine-tuned using the same labeled training and validation sets as the original optimal improved model to restore the performance lost during the pruning process.

6. The UAV small target detection method based on IASL-YOLO according to claim 5, characterized in that: The modules that are not allowed to be pruned include the Head part and the ASFF spatial adaptive network part of the model.

7. A UAV small target detection device based on IASL-YOLO, characterized in that: include: The first module is used to collect small target images of drones, and divide them into training set, verification set and test set after annotation; The second module is used to build an improved model consisting of Backbone, Neck and Head based on the YOLOv8s model. The Backbone part keeps the YOLOv8s model structure unchanged. The CFE-AFPN feature fusion module is introduced to replace the Neck part of YOLOv8s. The SIoU positioning loss function is introduced to the Head part to replace its original CIoU positioning loss function. The improved model is trained and tuned using labeled training sets and validation sets to obtain the optimal improved model. The third module is used to prune the optimal improved model generated in the training phase using a pruning algorithm to generate an 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