Unmanned aerial vehicle-oriented small target detection system and method based on YOLOv8s improved model, medium and equipment

Abstract

The invention provides an unmanned aerial vehicle-oriented small target detection system and method based on a YOLOv8s improved model, a medium and equipment, and relates to the field of computer vision, and the method comprises the steps: constructing a small target detection data set of an unmanned aerial vehicle image, and dividing the small target detection data set into a training set, a verification set and a test set; constructing a YOLOv8s network model for unmanned aerial vehicle image small target detection; inputting the training set and the verification set into a network model for training, generating a training weight file, introducing a WIoUv3 loss function, performing forward and backward propagation iterative optimization to finally obtain a trained network model, and performing evaluation by adopting a test set; and inputting a to-be-detected unmanned aerial vehicle image into the trained network model, and completing unmanned aerial vehicle image small target detection by adopting an SAHI reasoning strategy. According to the invention, the structure of the YOLOv8s network is improved, the perceptibility of the network to the small target is improved, and the detection precision of the small target in the unmanned aerial vehicle image is improved.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision, and in particular relates to a small target detection system, method, medium and device for unmanned aerial vehicle (UAV) based on an improved YOLOv8s model. Background Art

[0002] Small target detection has important practical application value in many fields. Usually, small targets refer to targets whose width or height is less than 10% of the image size, or whose area is less than a specific threshold (such as 32×32 pixels). With the rapid development of drone technology, especially driven by the continuous improvement of hardware performance, drone inspections have gradually replaced traditional manual inspections, showing many technical advantages. Drones can obtain large amounts of data, multi-perspective and ultra-high resolution image data in an efficient and flexible manner, providing important support for refined detection. However, compared with natural images, drone images have the characteristics of small target size and complex background, which brings great challenges to the task of small target detection. This particularity makes high-performance small target detection algorithms a key requirement, especially in the detection scenarios of static urban facilities such as manhole covers and traffic signs and dynamic targets such as vehicles and pedestrians. It is particularly important to develop algorithms that adapt to complex backgrounds and improve the accuracy of small target detection.

[0003] As a common target detection algorithm, YOLO has the advantages of fast detection speed, lightweight network structure, and convenient end-to-end training, and is widely used in various target detection tasks. However, the YOLO algorithm still has shortcomings in small target detection, such as the lack of sufficient consideration of small target features, which easily leads to feature loss, and limited target distinction ability in complex backgrounds, which is prone to false detection and missed detection. Summary of the invention

[0004] In view of the problems of low detection rate and high category judgment error rate in the existing YOLO algorithm in the application of UAV images, especially in the detection of small targets with complex backgrounds and fewer pixels, the present invention provides a small target detection system, method, medium and device for UAV based on the improved YOLOv8s model, so as to enhance the ability to extract small target features and reduce the interference of complex background on the detection results, thereby improving the accuracy and robustness of detection and meeting the detection requirements of high precision and multiple scenes.

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

[0006] In a first aspect, the present invention provides a small target detection system for drones based on an improved YOLOv8s model, comprising:

[0007] The dataset module is used to construct a small target detection dataset for drone images and divide it into training set, validation set and test set;

[0008] The network model module is used to build a YOLOv8s network model for small target detection in drone images, denoted as DSO-YOLOv8s network model;

[0009] The training module is used to input the training set and the validation set into the DSO-YOLOv8s network model for training, generate the training weight file, introduce the WIoUv3 loss function in the training, and finally obtain the trained DSO-YOLOv8s network model through forward and backward propagation iterative optimization, and use the test set for evaluation;

[0010] The detection module is used to input the drone image to be detected into the trained DSO-YOLOv8s network model and use the SAHI reasoning strategy to complete the small target detection in the drone image.

[0011] Optionally, in the network model module, the DSO-YOLOv8s network model includes a backbone network, a neck and a head;

[0012] The backbone network is constructed based on the CBS module, C2f module, C2F-BiFormer module and SPPF module, wherein the C2F-BiFormer module is a BiFormer module embedded in the C2f module, which is used to retain fine-grained context feature information during feature extraction; the backbone network outputs multiple feature maps of different levels according to the input drone image;

[0013] The neck adopts the PAFPN structure, performs multiple upsampling and feature concatenation operations on multiple feature maps of different levels output by the backbone network, and outputs multiple feature maps of different scales after the operation;

[0014] The head is provided with a plurality of detection heads of different scales, which respectively correspond to a plurality of feature maps of different scales output by the neck; the detection head is constructed using the C3Ghost module, and the features output by the neck are divided into a classification branch and a prediction box regression branch, which are respectively used to predict the target category and the prediction box position of the target.

[0015] Optionally, the backbone network includes a CBS module, a CBS module, a C2F-BiFormer module, a CBS module, a C2F-BiFormer module, a CBS module, a C2F-BiFormer module, a CBS module, a C2F module and an SPPF module connected in sequence, and a total of 4 feature maps of different levels are output at the three C2F-BiFormer modules and the SPPF module;

[0016] Among them, in the C2F-BiFormer module, the input feature map is divided into two paths, one path passes through the convolution layer and multiple BiFormer modules in sequence, and the other path only passes through the convolution layer, and then the two paths are merged for the Concat operation. The result after the Concat operation is then passed through the batch normalization layer, SiLU activation function and convolution layer in sequence before output.

[0017] Optionally, the neck fuses multi-scale features from feature maps at different levels through top-down and bottom-up paths to generate multiple feature maps of different scales.

[0018] Optionally, in the detection head of the head, for the input feature map The C3Ghost module first generates the main feature map through convolution operation :

[0019] ;

[0020] In the formula, Represents the weight of the convolution;

[0021] Then an additional ghost feature map is generated by linear transformation :

[0022] ;

[0023] In the formula, represents a linear transformation;

[0024] The final output features It is the main feature map And ghost feature map Combination of:

[0025] .

[0026] Optionally, in the training module, the calculation formula of the WIoUv3 loss function is as follows:

[0027] ;

[0028] ;

[0029] In the formula, is the WIoUv3 loss function, r is the non-monotonic focusing coefficient used to adjust the loss value, and is a tuning parameter used to adjust the quality cutoff criterion, represents the outlier degree of the anchor box, the anchor box represents the initial candidate box, Represents the initial WIoUv1 loss function, and the calculation formula is as follows:

[0030] ;

[0031] ;

[0032] ;

[0033] In the formula, is the intermediate parameter of the calculation, is the IoU loss function, represents the natural exponential function, and Represent the center coordinates of the prediction box and the target box respectively, and Represents the width and height of the minimum bounding rectangle of the prediction box and the target box, respectively. Indicates that and Separated from the feature map, IoU represents the intersection over union ratio of the predicted box and the target box. The predicted box is generated based on the anchor box and the predicted offset, and the target box is the real box marked in the dataset.

[0034] Optionally, the detection module inputs the drone image to be detected into the trained DSO-YOLOv8s network model, uses a fixed-size sliding window to segment the original image of the drone image, and the trained DSO-YOLOv8s network model performs target detection on each small image obtained after segmentation, and superimposes the prediction box on the original image; wherein, for multiple prediction boxes corresponding to each target, non-maximum suppression is used for filtering, and the best prediction box is retained.

[0035] In a second aspect, the present invention provides a method for detecting small targets of unmanned aerial vehicles based on an improved YOLOv8s model, comprising the following steps:

[0036] Step 1: Build a small target detection dataset of drone images and divide it into training set, validation set and test set;

[0037] Step 2: Build a YOLOv8s network model for small target detection in drone images, denoted as DSO-YOLOv8s network model;

[0038] Step 3: Input the training set and validation set into the DSO-YOLOv8s network model for training, generate a training weight file, introduce the WIoUv3 loss function in the training, optimize iteratively through forward and backward propagation, and finally obtain the trained DSO-YOLOv8s network model, and use the test set for evaluation;

[0039] Step 4: Input the drone image to be detected into the trained DSO-YOLOv8s network model, and use the SAHI reasoning strategy to complete the small target detection in the drone image.

[0040] In a third aspect, the present invention provides a computer-readable storage medium storing a computer program, wherein the computer program enables a computer to execute the small target detection method for drones based on the improved YOLOv8s model as described in the second aspect.

[0041] In a fourth aspect, the present invention provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method for detecting small targets for drones based on the improved YOLOv8s model as described in the second aspect is implemented.

[0042] The beneficial effects of the present invention are:

[0043] (1) To address the problem that the background of drone images is complex and small target detection is susceptible to noise interference, the present invention constructs a C2F-Biformer module based on the BiFormer module and the C2f module, and further constructs a backbone network based on the CBS module, the C2F-BiFormer module, the C2f module and the SPPF module. The idea of ​​capturing remote dependencies is introduced to retain fine-grained contextual feature information during feature extraction. At the same time, a total of 4 feature maps are output at the C2F-BiFormer module and the SPPF module to provide more location information of small targets.

[0044] (2) In order to improve the network's perception of small targets, the present invention performs three upsampling and three feature splicing operations on the neck, and fuses the high-level feature map with the position information features of the shallow network to make full use of the feature map; the present invention uses the C3Ghost module to construct the detection head, which reduces the model complexity and computational complexity while maintaining accuracy, and sets four detection heads of different scales on the head, which alleviates the problem of gradually decreasing resolution due to the large downsampling multiple in the YOLOv8s network, improves the model's positioning accuracy for small targets and maintains the integrity of semantic information.

[0045] (3) In order to alleviate the problem that the CIoU (Complete-IoU) loss is overly dependent on IoU (Intersection over Union) and center point distance, and easily ignores the need for precise positioning and shape matching in small target scenes, the present invention proposes to use WIoU (Wise Intersection over Union) v3 as the bounding box regression loss to reduce the situation where extreme samples produce excessive or harmful gradients, and pay more attention to samples of general quality, thereby improving the generalization ability and overall performance of the model.

[0046] (4) In order to improve the detection accuracy of small targets in the inference stage of the target detection algorithm, the present invention introduces the Slicing Aided Hyper Inference (SAHI) algorithm. This method creates a larger pixel area for small target objects in the image by slicing, retains the unique features of small targets and prevents feature loss, increases the detection effect of small targets in local areas, and effectively improves the overall detection performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a schematic diagram of the structure of the YOLOv8s network model for small target detection in drone images.

[0048] Figure 2 Schematic diagram of the structure of the C2F-BiFormer module.

[0049] Figure 3 It is a schematic diagram of the structure of the C2f module.

[0050] Figure 4 Schematic diagram of the structure of the BiFormer module.

[0051] Figure 5 Schematic diagram of the Bottleneck structure in the C2f module.

[0052] Figure 6 It is a structural diagram of the CBS module.

[0053] Figure 7 This is a structural comparison diagram of the traditional detection head and the improved small target detection head.

[0054] Figure 8 This is the schematic diagram of the C3Ghost module.

[0055] Fig. 9 The figure shows the visual comparison results of the predictions of the proposed method and traditional YOLOv8s on the VEDAI dataset.

[0056] Fig.10This is a visualization comparison result of the predictions of the proposed method and traditional YOLOv8s on the manhole cover target detection dataset. DETAILED DESCRIPTION

[0057] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0058] In one embodiment, the present invention proposes a method for detecting small targets for drones based on an improved YOLOv8s model, the method comprising the following steps:

[0059] S1: Construct a small target detection dataset for drone images and divide it into training set, validation set and test set.

[0060] This step specifically includes:

[0061] S11: Collect the VEDAI dataset for drone image vehicle detection, modify the annotation format to YOLO format through the algorithm, and divide the training set, validation set and test set into a ratio of 7:2:1 to construct the VEDAI target detection dataset.

[0062] S12: Collect high-resolution drone images, crop the drone images to generate 7323 sub-images of 512×512 pixels, and divide them into training set, validation set and test set in the ratio of 7:2:1. Use the LabelMe tool to annotate the manhole cover targets on the cropped samples and construct a drone image manhole cover target detection dataset.

[0063] S2: Build a YOLOv8s network model for small target detection in drone images (YOLOv8s for DroneSmall Object, DSO-YOLOv8s). The structure of the DSO-YOLOv8s network model is as follows: Figure 1 shown.

[0064] This step specifically includes:

[0065] S21: Construct a C2F-Biformer module and build a Backbone based on the CBS module, C2F-BiFormer module, C2f module and SPPF module. At the same time, a total of 4 feature maps are output at the C2F-BiFormer module and SPPF module to provide more location information of small targets.

[0066] Specifically, the C2F-BiFormer module is proposed to give full play to the efficient feature extraction of the C2f module and the global feature modeling ability of the BiFormer module, which can ensure that fine-grained contextual feature information is retained during the feature extraction process and improve the multi-scale perception and content perception capabilities of the target detection network. Figure 2 and Figure 3 As shown in Figure 1. The C2F-Biformer module is mainly composed of convolutional layers, BiFormer modules, Concat operations, and CBS modules. BiFormer is the basic building block of the BiFormer model. Figure 4 As shown in Figure 1. BiFormer uses Depthwise Separable Convolution (DWConv) to implicitly encode the relative position information of the input, and then uses Bi-Level Routing Attention (BRA) and Multilayer Perceptron (MLP) to model the cross-position relationship and position-by-position embedding of the input information, which has the ability to model global features. The Bottleneck structure in the C2f module and the CBS module structure are shown in Figure 1. Figure 5 and Figure 6 shown.

[0067] Compared with the original YOLOv8s, DSO-YOLOv8s provides shallower feature maps in the early stage of Backbone. Shallow feature maps have higher spatial resolution and contain richer local information. They are more sensitive to low-level features such as edges, textures, and colors. The finer edge and contour information and richer detail features make low-level feature maps helpful for segmentation and detection of small-scale targets.

[0068] S22: Perform upsampling and feature concatenation operations three times at the Neck end to fuse the high-level feature map with the location information features of the shallow network to make full use of the feature map, and set four detection heads of different scales at the Head end, such as Figure 7 shown.

[0069] Specifically, the three detection heads set by the traditional YOLOv8s have corresponding Neck feature maps of 64×64 for P3, 32×32 for P4, and 16×16 for P5, which are responsible for detecting small, medium, and large-scale targets, respectively. The downsampling multiple of YOLOv8s is relatively large, and the resolution gradually decreases, which will lead to the loss of manhole cover related information. It is difficult for the deeper feature maps to learn the feature information of small targets. The shallow feature maps have higher resolution and contain a large amount of location information of small targets.

[0070] In order to improve the positioning accuracy of the model for small targets while maintaining good semantic information, this embodiment performs three upsampling and three feature splicing operations on the Neck end to fuse the high-level feature map with the location information features of the shallow network. At the same time, four detection heads of different scales are set on the Head end to alleviate the problem of gradually decreasing resolution due to the large downsampling multiples in the YOLOv8s network, improve the positioning accuracy of the model for small targets and maintain the integrity of semantic information, so that the deep network can also learn the feature information of small targets and improve the detection effect of small targets.

[0071] S23: Use the C3Ghost module to rebuild the detection head, reducing the model complexity and computational complexity while maintaining accuracy, alleviating the problem of increased network parameters and computational complexity due to the introduction of large-size feature maps and multiple feature fusion operations.

[0072] like Figure 8 As shown, given the input feature map , the C3Ghost module first generates the main feature map through convolution operation :

[0073] ;

[0074] In the formula, Represents the weight of the convolution.

[0075] Then, additional ghost features are generated through linear transformation ,Right now:

[0076] ;

[0077] In the formula, Represents a linear transformation, such as a dot product or a weighted average.

[0078] The final output features It is the main feature map And ghost feature map Combination of:

[0079] .

[0080] This design greatly reduces the number of floating-point operations, while the fusion of the main feature map and the ghost feature map maintains the detection capability of the model, thereby achieving better real-time performance with fewer resources.

[0081] S3: Input the training set and validation set into the DSO-YOLOv8s network model for training, generate a training weight file, and introduce the WIoUv3 loss function. Through forward and backward propagation iterative optimization, the trained DSO-YOLOv8s network model is finally obtained.

[0082] This step specifically includes:

[0083] S31: Through input preprocessing (such as resizing and normalization), the image is adjusted to a fixed size of 512×512×3 for the adaptation model. Then enter Backbone and extract multi-scale features through each core module in turn. The core modules include CBS module, C2F-BiFormer module, C2f module and SPPF module. After passing through 3 C2F-Biformer modules and SPPF module, 4 efficient feature maps are output.

[0084] S32: Neck uses the PAFPN structure. The four feature maps output by Backbone fuse multi-scale features through top-down and bottom-up paths, and output four fused feature maps to further improve the detection capability of targets of different sizes.

[0085] S33: In the Head end, the features output by the Neck are divided into a classification branch and a prediction box regression branch by decoupling the detection head to predict the target category and the prediction box position respectively;

[0086] S34: Introduce the WIoUv3 loss function, optimize the model through the forward and backward propagation iterative algorithm, and finally obtain the trained DSO-YOLOv8s network model. Specifically, the use of the WIoUv3 loss function alleviates the excessive dependence of CIoU on IoU and center point distance, which easily ignores the problem of precise positioning and shape matching requirements in small target scenes. WIoUv3 introduces dynamic weighting in terms of center point offset, aspect ratio difference and IoU attenuation to make the matching of the predicted box and the target box more accurate. Among them, the anchor box refers to the initial candidate box generated by the network at a specific position, and the predicted box is the final prediction value generated based on the anchor box and the offset predicted by the network. The target box is the real box annotated in the data set, which is used to supervise the learning of the network. In addition, the introduction of the WIoUv3 loss function does not add additional parameters, keeping the model lightweight. The calculation formula of WIoUv3 is as follows:

[0087] ;

[0088] ;

[0089] In the formula, is the WIoUv3 loss function, is a non-monotonic focusing coefficient used to dynamically adjust the loss value. and is a tuning parameter that serves as a hyperparameter to tune the quality cutoff criterion, represents the outlier degree of the anchor box, It represents the initial WIoUv1 loss function, and the calculation formula is as follows:

[0090] ;

[0091] ;

[0092] ;

[0093] In the formula, represents the natural exponential function, and Represent the center coordinates of the prediction box and the target box respectively, and Represents the width and height of the minimum bounding rectangle of the prediction box and the target box, respectively. Indicates that and Separated from the feature map, IoU represents the intersection over union ratio between the predicted box and the target box. WIoUv3 uses a scaling factor right The dynamic weighting mechanism can adjust the weight according to the actual situation of the target, so as to perform better in small target detection, especially when dealing with overlapping target frames and dense targets.

[0094] S4: The test set is input into the trained DSO-YOLOv8s network model, and the SAHI reasoning strategy is introduced. The original image is segmented using a sliding window of a fixed size. Forward propagation target detection is applied to each overlapping slice to infer small targets. At the same time, full inference of the entire image is used to detect large targets. Finally, the predicted box is superimposed on the original image and filtered using non-maximum suppression (NMS). The predicted box that meets the IoU set threshold is matched, and the predicted box with low IoU is removed to finally generate a global detection result.

[0095] Cutting large-resolution images into smaller sub-images to avoid interference from large-scale background information can better focus on small targets and prevent them from being ignored in large images due to low resolution, thereby improving the network's attention to small targets.

[0096] S5: Based on the target detection results, the accuracy is evaluated using four indicators: precision, recall, mean average precision when intersection over union (IoU) is greater than 0.5 (mAP50), and mean average precision when IoU is between 0.5 and 0.95 (mAP50-95).

[0097] For the VEDAI target detection dataset, the mAP50 indicator is used for accuracy assessment; for the artificially constructed drone image manhole cover target detection dataset, the accuracy of four indicators including Precision, Recall, mAP50 and mAP50-95 is assessed. The calculation formulas for Precision and Recall are as follows:

[0098] ;

[0099] ;

[0100] In the formula, TP represents true positive examples, that is, the number of examples predicted to be in the current class and actually belonging to the current class; FP represents false positive examples, that is, the number of examples predicted to be in the current class but actually not belonging to the current class; FN represents false negative examples, that is, the number of examples actually belonging to the current class but predicted to be in other classes.

[0101] Average precision (AP) refers to the area under the Precision-Recall curve. mAP represents the average value of AP of all categories. The calculation formula is as follows:

[0102] ;

[0103] mAP50 and mAP50-95 are important indicators for evaluating the performance of target detection models, where mAP50 represents the mAP value when the IoU threshold is set to 0.5, and mAP50-95 represents the average mAP calculated under multiple IoU thresholds (from 0.5 to 0.95, with a step size of 0.05).

[0104] S6: In practical applications, the drone image to be detected is input into the trained DSO-YOLOv8s network model, and the SAHI reasoning strategy is used to complete the small target detection in the drone image.

[0105] Next, this embodiment conducts experiments on two remote sensing image small target detection datasets. The data used are the VEDAI target detection dataset for drone image vehicle detection and the artificially constructed drone image manhole cover target detection dataset. The vehicles and manhole cover targets in the data meet the definition of small targets. Among them, the VEDAI target detection dataset is a public dataset for remote sensing image vehicle detection, containing about 1210 images with a resolution of 512×512 or 1024×1024, annotating 9 vehicle categories (such as cars, trucks, airplanes, etc.) and more than 2950 targets, and providing two annotation formats: rotating bounding box and rectangular bounding box. The dataset covers a variety of scenes such as cities, villages, and mountains, and has the characteristics of changing illumination, complex background, and diverse target scales. It is suitable for studying target detection, multi-scale detection, and target detection algorithms. However, the number of targets in each image in this dataset is small, and the background contains a variety of geographical features (buildings, vegetation, roads, etc.), which increases the difficulty of detection. The proportion of vehicles in the image area is very small, which puts higher requirements on the detection model. The manhole cover target detection dataset comes from orthophoto results with a spatial resolution of 3cm. After cropping, there are 7323 sub-images of 512×512 pixels, which are divided into training set, validation set and test set in the ratio of 7:2:1 and annotated using LabelMe.

[0106] In order to verify the performance of the DSO-YOLOv8s network model, this embodiment selects the following methods for comparative experiments. Table 1 shows the hyperparameters and hardware device information in the experiment.

[0107] Table 1 Experimental parameters

[0108] Table 2 shows the comparison of detection accuracy and model parameter quantity of each comparison method on the VEDAI target detection dataset, and the visual comparison of some detection results of YOLOv8s is shown in the figure below. Fig. 9As shown in the table, DSO-YOLOv8s achieved the highest mAP50 of 62.9%, which is 8.1% higher than YOLOv5s (54.8%), and also improved on YOLOv4 (62.4%), YOLOv3 (61.1%) and YOLOv8s (59.2%), which increased by 0.5%, 1.8% and 3.7% respectively. In terms of single-category detection performance, DSO-YOLOv8s performed well in certain specific categories, especially in the Boat and Van categories, achieving the highest mAP50 values ​​of 53.5% and 67.4% respectively, leading other models, and maintaining high accuracy in multiple other categories, which shows that DSO-YOLOv8s has a relatively balanced detection accuracy in various categories. In terms of model complexity, the number of parameters of DSO-YOLOv8s is 10.9M, which is much lower than YOLOv3 (61.54M) and YOLOv4 (52.51M), slightly lower than YOLOv8s's 11.2M, and only slightly higher than YOLOv5s (7.07M). This lightweight design enables the model to significantly reduce computational costs while ensuring high accuracy. It can also be seen from the visual comparison chart that compared with YOLOv8s, DSO-YOLOv8s can detect some small targets that are ignored by YOLOv8s, showing better detection capabilities. In summary, DSO-YOLOv8s demonstrated excellent overall performance in this comparative experiment. It not only outperformed the traditional YOLO series models in terms of overall mAP50 and mAP50 of multiple categories, but also effectively optimized the number of parameters and computational requirements, achieving high computational efficiency and resource utilization. Especially in specific categories such as Boat and Van, it shows significant advantages, demonstrating its strong potential in diverse object detection tasks.

[0109] Table 2 Detection accuracy and performance indicators of different models on the VEDAI dataset

[0110] Table 3 shows the detection accuracy and performance indicators of each comparison method on the manhole cover test set, and the visual comparison of the detection results with YOLOv8s is shown in the figure below. Fig.10As shown in Table 3, compared with YOLO's traditional models (YOLOv3-tiny, YOLOv5s, YOLOv6s, YOLOv8s and YOLOv10s), DSO-YOLOv8s shows significant advantages in detection performance. Specifically, the Precision of DSO-YOLOv8s reaches 78.2%, which is 3.8% higher than the best performing YOLOv10s. Compared with YOLOv3-tiny, YOLOv5s, YOLOv6s and YOLOv8s, it is improved by 9.6%, 5.9%, 4.4% and 5.7% respectively, reflecting better detection accuracy. In addition, DSO-YOLOv8s also surpasses traditional models in Recall, mAP50 and mAP50-95, which are 69.5%, 77.9% and 40.1% respectively. Compared with YOLOv3-tiny, YOLOv5s, YOLOv6s, YOLOv8s and YOLOv10s, mAP50 is improved by 12.5%, 5.1%, 5.1%, 4.0% and 5.3% respectively.

[0111] Compared with other improved models (such as Faster-RCNN, Retinanet, GFL and ARSD), DSO-YOLOv8s also has obvious advantages in performance. In particular, compared with ARSD, which has relatively high performance, DSO-YOLOv8s's Precision is improved by 5.1%, Recall is improved by 1.3%, and mAP50 and mAP50-95 are improved by 4.6% and 2.5% respectively.

[0112] At the same time, the number of parameters of DSO-YOLOv8s is 8.2M, which is 3.0M lower than that of YOLOv8s, and significantly reduced compared with models such as Retinanet and Faster-RCNN. It is only 1 / 7 and 1 / 5 of Retinanet (55.4M) and Faster-RCNN (41.3M), and only slightly higher than YOLOv5s and YOLOv10s, showing higher model efficiency.

[0113] Table 3 Detection accuracy and performance indicators of different models on the manhole cover test set

[0114] It can also be seen from the visual comparison chart that compared with YOLOv8s, DSO-YOLOv8s can detect some manhole covers that are ignored by YOLOv8s, showing better detection capabilities. DSO-YOLOv8s is significantly better than the traditional YOLO model and other improved models in terms of performance indicators such as Precision and Recall, while maintaining a low number of parameters. This shows that DSO-YOLOv8s achieves high-precision detection while taking into account the lightweight design of the model, and has great potential for practical applications.

[0115] In another embodiment, the present invention proposes a small target detection system for unmanned aerial vehicles based on an improved YOLOv8s model, which corresponds to the small target detection method for unmanned aerial vehicles based on an improved YOLOv8s model in the aforementioned embodiment, and specifically includes the following modules:

[0116] The dataset module is used to construct a small target detection dataset for drone images and divide it into training set, validation set and test set;

[0117] The network model module is used to build a YOLOv8s network model for small target detection in drone images, denoted as DSO-YOLOv8s network model;

[0118] The training module is used to input the training set and the validation set into the DSO-YOLOv8s network model for training, generate the training weight file, introduce the WIoUv3 loss function in the training, and finally obtain the trained DSO-YOLOv8s network model through forward and backward propagation iterative optimization, and use the test set for evaluation;

[0119] The detection module is used to input the drone image to be detected into the trained DSO-YOLOv8s network model and use the SAHI reasoning strategy to complete the small target detection in the drone image.

[0120] In this embodiment, the functions and execution processes of each module of the system correspond to the steps of the small target detection method for drones based on the improved YOLOv8s model in the aforementioned embodiment, so they are not repeated here.

[0121] In another embodiment, the present invention proposes a computer-readable storage medium storing a computer program, which enables a computer to execute the small target detection method for drones based on the improved YOLOv8s model of the aforementioned embodiment.

[0122] In another embodiment, the present invention proposes an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the small target detection method for drones based on the improved YOLOv8s model of the aforementioned embodiment is implemented.

[0123] In the embodiments disclosed in the present application, the computer storage medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or apparatus. The computer storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or equipment, or any suitable combination of the foregoing. More specific examples of computer storage media may include electrical connections based on one or more lines, portable computer disks, hard disks, random access memories (RAM), read-only memories (ROM), erasable programmable read-only memories (EPROM or flash memory), optical fibers, portable compact disk read-only memories (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0124] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0125] The above are only preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should be regarded as the protection scope of the present invention.

Claims

1. A small target detection system for drones based on the improved YOLOv8s model, characterized by: Including dataset module, network model module, training module and detection module: The dataset module is used to construct a small target detection dataset for drone images and divide it into training set, validation set and test set; The network model module is used to build a YOLOv8s network model for small target detection in drone images, denoted as DSO-YOLOv8s network model; The training module is used to input the training set and the validation set into the DSO-YOLOv8s network model for training, generate the training weight file, introduce the WIoUv3 loss function in the training, and finally obtain the trained DSO-YOLOv8s network model through forward and backward propagation iterative optimization, and use the test set for evaluation; The detection module is used to input the drone image to be detected into the trained DSO-YOLOv8s network model and use the SAHI reasoning strategy to complete the small target detection in the drone image.

2. The UAV-oriented small target detection system based on the improved YOLOv8s model as claimed in claim 1, characterized in that: In the network model module, the DSO-YOLOv8s network model includes a backbone network, a neck and a head; The backbone network is constructed based on the CBS module, C2f module, C2F-BiFormer module and SPPF module, wherein the C2F-BiFormer module is a BiFormer module embedded in the C2f module, which is used to retain fine-grained context feature information during feature extraction; the backbone network outputs multiple feature maps of different levels according to the input drone image; The neck adopts the PAFPN structure, performs multiple upsampling and feature concatenation operations on multiple feature maps of different levels output by the backbone network, and outputs multiple feature maps of different scales after the operation; The head is provided with a plurality of detection heads of different scales, which respectively correspond to a plurality of feature maps of different scales output by the neck; the detection head is constructed using the C3Ghost module, and the features output by the neck are divided into a classification branch and a prediction box regression branch, which are respectively used to predict the target category and the prediction box position of the target.

3. The UAV-oriented small target detection system based on the improved YOLOv8s model as claimed in claim 2, characterized in that: The backbone network includes a CBS module, a CBS module, a C2F-BiFormer module, a CBS module, a C2F-BiFormer module, a CBS module, a C2F-BiFormer module, a CBS module, a C2F module and an SPPF module connected in sequence, and outputs a total of 4 feature maps of different levels at the three C2F-BiFormer modules and the SPPF module; Among them, in the C2F-BiFormer module, the input feature map is divided into two paths, one path passes through the convolution layer and multiple BiFormer modules in sequence, and the other path only passes through the convolution layer, and then the two paths are merged for the Concat operation. The result after the Concat operation is then passed through the batch normalization layer, SiLU activation function and convolution layer in sequence before output.

4. The UAV-oriented small target detection system based on the improved YOLOv8s model as claimed in claim 2, characterized in that: The neck fuses multi-scale features from feature maps at different levels through top-down and bottom-up paths to generate multiple feature maps of different scales.

5. The UAV-oriented small target detection system based on the improved YOLOv8s model as claimed in claim 2, characterized in that: In the detection head of the head, for the input feature map The C3Ghost module first generates the main feature map through convolution operation : ； In the formula, Represents the weight of the convolution; Then an additional ghost feature map is generated by linear transformation : ； In the formula, represents a linear transformation; The final output features It is the main feature map And ghost feature map Combination of: 。 6. The UAV-oriented small target detection system based on the improved YOLOv8s model as claimed in claim 1, characterized in that: In the training module, the calculation formula of the WIoUv3 loss function is as follows: ； ； In the formula, is the WIoUv3 loss function, r is the non-monotonic focusing coefficient used to adjust the loss value, and is a tuning parameter used to adjust the quality cutoff criterion, represents the outlier degree of the anchor box, the anchor box represents the initial candidate box, Represents the initial WIoUv1 loss function, and the calculation formula is as follows: ； ； ； In the formula, is the intermediate parameter of the calculation, is the IoU loss function, represents the natural exponential function, and Represent the center coordinates of the prediction box and the target box respectively, and Represents the width and height of the minimum bounding rectangle of the prediction box and the target box, respectively. Indicates that and Separated from the feature map, IoU represents the intersection over union ratio of the predicted box and the target box. The predicted box is generated based on the anchor box and the predicted offset, and the target box is the real box marked in the dataset.

7. The UAV-oriented small target detection system based on the improved YOLOv8s model as claimed in claim 1, characterized in that: The detection module inputs the drone image to be detected into the trained DSO-YOLOv8s network model, uses a fixed-size sliding window to segment the original image of the drone image, and the trained DSO-YOLOv8s network model performs target detection on each small image obtained after segmentation, and superimposes the prediction frame on the original image; wherein, for multiple prediction frames corresponding to each target, non-maximum suppression is used for filtering, and the best prediction frame is retained.

8. A small target detection method for drones based on the improved YOLOv8s model, characterized in that: The steps include: Step 1: Build a small target detection dataset for drone images and divide it into training set, validation set and test set; Step 2: Build a YOLOv8s network model for small target detection in drone images, denoted as DSO-YOLOv8s network model; Step 3: Input the training set and validation set into the DSO-YOLOv8s network model for training, generate a training weight file, introduce the WIoUv3 loss function in the training, optimize iteratively through forward and backward propagation, and finally obtain the trained DSO-YOLOv8s network model, and use the test set for evaluation; Step 4: Input the drone image to be detected into the trained DSO-YOLOv8s network model, and use the SAHI reasoning strategy to complete the small target detection in the drone image.

9. A computer-readable storage medium storing a computer program, characterized in that: The computer program enables the computer to execute the small target detection method for drones based on the improved YOLOv8s model as described in claim 8.

10. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method for detecting small targets for drones based on the improved YOLOv8s model as claimed in claim 8 is implemented.

Images