Improved SAR image multi-target detection method based on YOLOv8

Abstract

The invention discloses an improved SAR (Synthetic Aperture Radar) image multi-target detection method based on YOLOv8, which belongs to the technical field of SAR image recognition, and comprises the following steps: carrying out improvement processing on a YOLOv8 model; training the improved YOLOv8 model by adopting a public SAR image data set to obtain a lightweight SAR image detection model; a lightweight SAR image detection model is adopted to identify multiple targets in the SAR image; and performing performance evaluation on the lightweight SAR image detection model according to a multi-target detection result. According to the invention, model improvement is carried out based on YOLOv8, and the improved YOLOv8 model has high precision for SAR image multi-target identification and is suitable for SAR image multi-class target detection tasks.

Description

Technical Field

[0001] The present invention belongs to the technical field of SAR image recognition, and in particular relates to a SAR image multi-target detection method based on an improved YOLOv8. Background Art

[0002] Synthetic Aperture Radar (SAR) is an active microwave remote sensing technology that generates high-resolution images by moving the antenna at a constant speed along a long linear array track, emitting coherent signals, and coherently processing the echoes received at different positions. Compared with traditional optical sensors, SAR is not restricted by factors such as light intensity and weather when imaging or detecting, and has the advantages of all-weather and all-day. The emergence of SAR imaging technology makes up for the defect that optical imaging technology cannot map the surface at night and in bad weather conditions. Among them, in the civilian field, SAR is widely used in mineral resource exploration, disaster monitoring and prevention, topographic mapping, as well as ocean, agriculture, forestry and other aspects, and plays an important role. Due to the unique performance of SAR, countries attach great importance to the interpretation technology of typical targets and continue to promote its development and progress. In the real-time detection of marine and land targets, especially the identification of ships and aircraft, SAR images show obvious advantages. However, the existing deep learning detection algorithm still has a lot of room for improvement under the influence of complex background and noise, so it is necessary to detect ships and aircraft. On the sea, timely monitoring of ships is helpful for marine safety, shipping management, prevention and control of illegal activities and safeguarding national territorial integrity. On land, identifying the number and location of aircraft is of great value for airport management and military intelligence, and can effectively improve safety and efficiency. At the same time, with the development of deep learning, the performance of SAR image target detection has been greatly improved. At present, the SAR image target detection algorithm based on deep learning still has a lot of room for improvement in detection speed, detection accuracy and other aspects.

[0003] Object detection algorithms based on deep learning are mainly divided into two categories: one is the two-stage detection algorithm represented by regional convolutional neural network R-CNN, Fast R-CNN, Faster R-CNN and DETR, and the other is the single-stage detection algorithm represented by YOLO, SSD, Retina-Net and Fcos. Among them, the two-stage algorithm first distinguishes the foreground and background through the region proposal network (RPN), and then classifies and predicts the position of the target. Although this type of algorithm has high detection accuracy, the detection speed is relatively slow. In contrast, the single-stage algorithm does not need to generate candidate regions, but simplifies the object detection problem into a regression problem, but the detection speed is significantly faster than the two-stage algorithm, so it is more suitable for real-time detection tasks.

[0004] Although ship target detection in SAR images based on deep learning has achieved relatively good detection results, there are still great challenges for small targets in the scene and targets with complex backgrounds. First, the size and shape of targets vary. Coupled with factors such as climate, viewing angle, lighting, and noise generated by coherent imaging, target detection is prone to false detection and missed detection. Secondly, there are a lot of other backgrounds in the image, which makes it difficult for the network to distinguish between the target and the background. In addition, although lightweight networks are emerging in an endless stream, how to improve accuracy while ensuring that too much additional computation is not added cannot be ignored. Summary of the invention

[0005] The purpose of the present invention is to provide a SAR image multi-target detection method based on YOLOv8 to address the above-mentioned deficiencies in the prior art, so as to solve the problem of low detection accuracy for small targets in the scene and targets with complex backgrounds in the existing SAR image target detection.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is:

[0007] A SAR image multi-target detection method based on YOLOv8 improvement includes the following steps:

[0008] S1. Improve the YOLOv8 model;

[0009] S2. Use the public SAR image dataset to train the improved YOLOv8 model to obtain a lightweight SAR image detection model;

[0010] S3, using lightweight SAR image detection model to identify multiple targets in SAR images;

[0011] S4. Perform performance evaluation on lightweight SAR image detection model based on multi-target detection results.

[0012] Furthermore, S1 includes the following sub-steps:

[0013] S11. Add the global maximum pooling path and the global average pooling path to the SPPF module in the YOLOv8 model backbone network, and add the CA coordinate attention mechanism after the output of the SPPF_improve module;

[0014] S12, use dynamic upsampling DySample to replace the upsample Upsample in the YOLOv8 model;

[0015] S13, use the dilated residual module DWR to replace the Bottleneck module in C2f in the YOLOv8 model;

[0016] S14. Modify the CIoU bounding box loss function in the YOLOv8 model to the Inner-CIoU loss function. Further, in S14, the Inner-CIoU loss function is:

[0017] L Inner-CoU =L CIoU +IoU-IoU inner

[0018] Where, L Inner-CoU is the Inner-CIoU loss function; L CIoU is the original CIoU loss function; IoU is the intersection-over-union ratio between the original predicted box and the real box; IoU inner is the embedded IoU, which represents the intersection over union (IoU) of the predicted box and the true box after the scaling factor is introduced.

[0019] Furthermore, the original CIoU loss function L CIoU for:

[0020]

[0021]

[0022]

[0023]

[0024] Where B is the area of ​​the prediction box; B gt is the area of ​​the real box; υ is a parameter to measure the consistency of aspect ratio; w gt is the width of the real frame; h gt is the height of the real box; w is the width of the predicted box; h is the height of the predicted box; α is the balance factor; ρ is the Euclidean distance between the predicted box and the center point of the real box; b is the coordinate of the center point of the predicted box; b gt is the coordinate of the center point of the real box; c is the diagonal length of the minimum circumscribed rectangle of the predicted box and the real box.

[0025] Furthermore, based on the original CIoU loss function L CIoU , introduce scaling auxiliary bounding box ratio and calculate IoU inner for:

[0026]

[0027]

[0028] In the formula, are the left, right, top and bottom boundaries of the ground-truth box respectively; They are the horizontal and vertical coordinates of the center point of the real frame respectively;

[0029] according to And the four values ​​​​b of the prediction box l 、b r 、b t and b b , calculate the intersection inter of the scaled prediction box and the true box and the union union of the scaled prediction box and the true box:

[0030]

[0031] union=(w gt ×h gt )×ratio 2 +(w×h)×ratio 2 -inter

[0032] Calculate IoU based on inter and union inner for:

[0033]

[0034] Where inter is the intersection of the scaled prediction box and the true box; union is the union of the scaled prediction box and the true box.

[0035] Furthermore, the SAR image dataset in S2 is the public MSAR-1.0 dataset, which is randomly divided into training set, validation set and test set in a ratio of 7:1:2.

[0036] Furthermore, the test set is used in S4 to evaluate the performance of the lightweight SAR image detection model:

[0037]

[0038]

[0039] AP = ∫0 1 P(R)dR

[0040]

[0041] Where Precision is the accuracy rate; TP is the number of positive classes predicted as positive classes; FP is the number of negative classes predicted as positive classes; FN is the number of positive classes predicted as negative classes; Recall is the recall rate; AP is the precision rate; mAP is the average precision rate; P(R) is the relationship curve between precision and recall rate with horizontal and vertical coordinates; R is the recall rate; n is the number of categories of the detected targets; AP is the average precision rate; i is the average precision of the ith category.

[0042] The improved SAR image multi-target detection method based on YOLOv8 provided by the present invention has the following beneficial effects:

[0043] 1. The present invention improves the model based on YOLOv8. The improved YOLOv8 model has higher accuracy for multi-target recognition in SAR images and is suitable for multi-category target detection tasks in SAR images.

[0044] 2. SPPF module enhancement: Global maximum pooling and global average pooling paths are added to the original SPPF module of YOLOv8, so that the network can more effectively capture multi-scale contextual information. Maximum pooling emphasizes prominent edge features, and average pooling reduces the impact of noise, thereby improving the diversity of feature expression.

[0045] 3. CA Coordinate Attention Mechanism: The CA (Coordinate Attention) mechanism is introduced to integrate position information into channel attention, so that the model can focus on the different position features of the target more accurately, which helps to achieve more accurate detection in complex backgrounds.

[0046] 4. DySample dynamic upsampling: The DySample module is used to replace the traditional upsampling module. Through adaptive sampling and dynamic offset, the model can better restore details during upsampling and adapt to different input features. This method enhances the content perception ability of feature maps and improves the performance of detection tasks in complex scenes and high-resolution output.

[0047] 5. DWR dilated residual module: The DWR (Dilation-wise Residual) dilated residual module is introduced in the C2f module of the backbone network. The multi-scale receptive field is realized through convolution operations with different dilation rates, thereby enhancing the feature extraction capabilities of large-scale and small-scale targets.

[0048] 6. Inner-CIoU loss function: Replace the CIoU loss function with Inner-CIoU to enhance the model's adaptability to samples with different IoUs and improve the convergence speed and stability of bounding box prediction. Inner-CIoU scales the auxiliary bounding box so that high IoU samples converge more accurately when they are close to the true boundary, and low IoU samples have wider coverage when the network adjusts the bounding box. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a flow chart of the improved SAR image multi-target detection method based on YOLOv8 in the present invention.

[0050] Figure 2 This is a diagram of the improved YOLOv8 network structure of the present invention.

[0051] Figure 3 This is the structural diagram of the SPPF_improve module of the present invention.

[0052] Figure 4 This is the structural diagram of the CA attention mechanism module of the present invention.

[0053] Figure 5 This is a structural diagram of the Dysample module of the present invention.

[0054] Figure 6 This is a structural diagram of the DWR and C2f_DWR modules of the present invention.

[0055] Figure 7 This is a graph of the annotation results of the original data set of the present invention.

[0056] Figure 8 This is the original YOLOv8 detection result diagram of the present invention.

[0057] Fig. 9 This is a diagram of the improved YOLOv8 detection results shown in the present invention. DETAILED DESCRIPTION

[0058] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0059] This embodiment provides a SAR image multi-target detection method based on YOLOv8. This embodiment has high accuracy for multi-target recognition of small targets in the scene and complex backgrounds without adding too much additional calculation. Figure 1 , which specifically includes the following steps:

[0060] Step S1, improving the YOLOv8 model;

[0061] refer to Figure 2 In this embodiment, the YOLOv8 model is light-weighted and improved. The improved target detection network includes a backbone feature extraction module (Backbone), a feature fusion module (Neck), and a target prediction module (Head) connected from left to right, which includes the following steps:

[0062] Step S11, add a global maximum pooling path and a global average pooling path in the SPPF module in the YOLOv8 model backbone network, and add a CA coordinate attention mechanism after the output of the SPPF_improve module;

[0063] Specifically, the SPPF module in the backbone network adds two additional paths: global maximum pooling and global average pooling, which reduces the number of input channels by half, reduces the computational cost of the previous convolution operation, and enables the network to obtain more contextual information from different scales, further enriching the feature expression. Global maximum pooling is better at retaining edges and prominent features, while global average pooling helps reduce the impact of noise and increases feature diversity. A CA coordinate attention mechanism is added after the SPPF_improve module. The CA coordinate attention mechanism can effectively capture the position information of the input feature map. Unlike traditional channel attention, CA introduces coordinate information, which not only considers information between channels, but also position information, allowing the model to accurately enhance features at different positions. This is especially important for the detection and recognition of targets in complex backgrounds. The SPPF_improve module is as follows Figure 3 As shown, the CA coordinate attention mechanism is as follows Figure 4 shown.

[0064] Step S12, using dynamic upsampling DySample to replace the upsample Upsample in the YOLOv8 model;

[0065] After replacement, this embodiment can adaptively sample and dynamically offset the mechanism, which enables the model to better restore details, enhance information in feature maps, and more flexibly handle different input features during upsampling. For complex scenes, tasks with significant scale changes, or applications that require high-resolution output, DySample can significantly improve the expressiveness of the model, and all this is achieved at a relatively reasonable computational complexity.

[0066] Specifically, DySample uses an efficient point sampling method to replace the traditional dynamic convolution kernel, while enhancing the content perception of the feature map. This design not only greatly reduces the computational overhead, but also enhances the model's ability to perceive the content of the feature map, making DySample perform well in multiple intensive prediction tasks (such as semantic segmentation, object detection, and instance segmentation). The DySampel module is as follows Figure 5 Middle: Input features, generated offsets, and original mesh are denoted by χ, o, and g, respectively. Figure 5 The above figure is a fixed range sampling method. The offset and range are relatively simple, which is suitable for scenes with less local feature changes. The feature map input in the figure below is (X). Similarly, the input is first linearly transformed to H×W×2gs 2size. In this structure, the range factor is generated dynamically, that is, an adaptive range factor is generated according to the specific situation of the input features. This dynamic range factor allows the offset range to be dynamically adjusted according to the features of different inputs. Then the pixel shuffle is the same as the static range factor. Pixel Shuffle is used to rearrange the feature map and adjust the spatial dimension. The g module is the original sampling network, which is responsible for sampling the input feature map and generating basic feature extraction. The o module is responsible for generating the offset, which is used to fine-tune the sampling of the feature map. By combining the fixed range factor, the change of the offset is constrained. DySample reduces inference time, memory usage, and computation while maintaining high performance, making it a general and efficient dynamic upsampling solution.

[0067] Step S13, using the dilated residual module DWR to replace the Bottleneck module in C2f in the YOLOv8 model, which can be used to extract high-level network features, wherein the multi-branch structure is used to expand the receptive field, and each branch uses a hole depth convolution with a different hole rate;

[0068] Specifically, the DWR modules are as follows Figure 6 As shown on the left, convolution operations with different expansion rates d (1, 3, 5) are used to capture features of different scales. Finally, through feature concatenation and convolution fusion, the design of residual connections helps information flow and maintains the depth and performance of the model. In SAR images, the scale and shape of targets may vary greatly, especially in complex scenes, where small and large targets often coexist. Through this multi-scale receptive field design, the DWR module can more comprehensively capture target features at different scales. This is more advantageous than the original Bottleneck module in that:

[0069] 1. Capture the global features of large-scale targets;

[0070] 2. For small targets, a larger receptive field can be used to enhance local features.

[0071] Although the improvement of the C2f module has led to an increase in the number of model parameters, its ability to extract target features has been significantly enhanced, especially in complex background scenes. The model performance is more robust, the detection accuracy is significantly improved, and the increase in inference speed has no significant impact on real-time performance.

[0072] Step S14, modify the CIoU bounding box loss function in the YOLOv8 model to the Inner-CIoU loss function; the Inner-CIoU loss function is improved from the original CIoU bounding box loss function, and introduces a scaling auxiliary box. By introducing a scaling ratio (for example, ratio = 0.7) around the target bounding box, the embedded IoU of the target is calculated. Among them, Inner-CIoU introduces an auxiliary box. By scaling the auxiliary bounding box, not only can the model's adaptability to different IoU samples be enhanced, but also the convergence speed can be improved during training. For samples with high IoU, smaller auxiliary boxes can help the model converge more accurately when approaching the true boundary. For samples with low IoU, larger auxiliary boxes can cover more potential positions, allowing the network to better adjust the bounding box. This loss function improves the adaptability, convergence speed, and gradient adjustment ability of the IoU loss.

[0073] This embodiment has the following advantages:

[0074] Enhanced convergence, more stable bounding box regression, improved small target detection capabilities, reduced degradation and compatible with existing IoU improved algorithms, can be combined with existing loss functions. The present invention uses the original CIoU for combination. In general, the design focus of Inner-CIoU is to make bounding box prediction more accurate by paying more attention to the matching of the core area inside the target, thereby improving the robustness and performance in the detection task.

[0075] Specifically, the original CIoU loss function L CIoU The calculation formula is as follows:

[0076]

[0077]

[0078]

[0079] Where B is the area of ​​the prediction box; B gt is the area of ​​the real box; υ is a parameter to measure the consistency of aspect ratio; w gt is the width of the real frame; h gt is the height of the real box; w is the width of the predicted box; h is the height of the predicted box; α is the balance factor; ρ is the Euclidean distance between the predicted box and the center point of the real box; b is the coordinate of the center point of the predicted box; b gt is the coordinate of the center point of the real box; c is the diagonal length of the minimum circumscribed rectangle of the predicted box and the real box.

[0080] The above is the original CIoU loss function. On this basis, the scaling auxiliary bounding box is introduced with ratio = 0.7 (generally recommended between 0.5 and 1.5), which is adjusted according to the specific features of the target to be detected) as follows:

[0081]

[0082]

[0083] In the formula, are the left, right, top and bottom boundaries of the ground-truth box respectively; They are the horizontal and vertical coordinates of the center point of the real frame respectively;

[0084] The above four formulas are obtained by multiplying the ground truth by the auxiliary box scaling factor ratio. Similarly, the predicted box has the same four values, namely b l 、b r 、b t and b b , then calculate inter and union, and finally get IoU inner The formula is as follows:

[0085]

[0086] union=(w gt ×h gt )×ratio 2 +(w×h)×ratio 2 -inter

[0087]

[0088] The Inner-CIoU loss function calculates the auxiliary Inner-IoU based on the original CIoU loss by introducing the auxiliary box (multiplied by the scaling factor ratio), thereby improving the adaptability of the loss function. The final form of the loss function is the original CIoU loss and IoU minus the auxiliary box IoU inner The sum of the differences:

[0089] L Inner-CoU =L CIoU +IoU-IoU inner

[0090] Where, L Inner-CoU is the Inner-CIoU loss function; L CIoU is the original CIoU loss function; IoU is the intersection-over-union ratio between the original predicted box and the real box; IoU inner It is the embedded IoU, the intersection over union (IoU) of the predicted box and the true box after introducing the scaling factor.

[0091] Step S2: using a public SAR image dataset to train an improved YOLOv8 model to obtain a lightweight SAR image detection model;

[0092] The original data set of MSAR-1.0 is input into the initial YOLOv8 network for training, and the model weight file obtained is used as the baseline network. The improved YOLOv8 model is trained multiple times to obtain the optimal result as the final lightweight SAR image detection model.

[0093] Specifically, the data set of this embodiment is a public MSAR-1.0 data set, which contains 28,449 detection slices, and the data comes from the Haisi-1 satellite and the Gaofen-3 satellite. The scenes covered by this data set include airports, ports, nearshore, islands, offshore, urban areas, etc. The target types include four categories: aircraft, oil tanks, bridges and ships, with a total of 1,851 bridges, 39,858 ships, 12,319 oil tanks and 6,368 aircraft. The data set is randomly divided into a training set, a validation set and a test set in a ratio of 7:1:2. The training set is used for training the initial target network, the validation set is used to evaluate the model performance and adjust the hyperparameters during the training process, and the test set is used to finally evaluate the generalization ability of the model on unseen data.

[0094] During specific training, the system environment used for training in the embodiment of the present invention is Windows 10, the graphics card model is NVIDIA TITAN Xp (12GB), and the training and testing environment is built based on CUDA 12.1 and Pytorch 2.5.0 framework.

[0095] Set the required parameters, including epoch (iteration number), batchsize (number of images in each batch), lr0 (initial learning rate), lrf (final learning rate), momentum (momentum parameter), etc. epoch is 200, batchsize is 16, lr0 is 0.01, lrf is 0.002, momentum is 0.9. The iou threshold is 0.5, and the SGD optimizer is used to iteratively update the network model parameters.

[0096] Step S3: Reference Figure 7 to Figure 9 , a lightweight SAR image detection model is used to identify multiple targets in SAR images;

[0097] This embodiment uses a lightweight SAR image detection model to identify four types of targets in SAR images: airplanes, oil tanks, bridges, and ships.

[0098] Step S4: perform performance evaluation on the lightweight SAR image detection model according to the multi-target detection results. The performance evaluation of the model in this embodiment is mainly carried out from two aspects: detection accuracy and model complexity.

[0099] The evaluation indicators of detection accuracy include precision, recall, mean average precision (mAP) and inference speed. The model complexity includes model size and number of parameters.

[0100] When defining the indicators, TP represents the number of positive classes predicted as positive classes, FP represents the number of negative classes predicted as positive classes, and FN represents the number of positive classes predicted as negative classes. Based on this, the calculation formulas for precision, recall, and mean average precision (mAP) are as follows:

[0101]

[0102]

[0103] AP = ∫0 1 P(R)dR

[0104]

[0105] Where P(R) is the relationship curve between precision and recall, R is the recall rate, n is the number of categories of the detected targets, the detection categories of this method are airplanes, oil tanks, bridges and ships, so n is 4; AP i is the average precision of the ith category.

[0106] Based on the same SAR image data set, the original YOLOv8 model (scale selection YOLOv8n) and the method of the present invention (YOLOv8 improvement) are trained, and the target objects are set as airplanes, oil tanks, bridges, and ships. The trained target detection network is evaluated as above, so as to obtain the results shown in Table 1.

[0107] Table 1 Comparative experimental results

[0108]

[0109]

[0110] As shown in Table 1 above, the detection effect of the method of the present invention is better than that of YOLOv8n. The overall mAP@0.5 index of the four types of targets, namely airplane, oil tank, bridge and ship, has increased from 0.899 to 0.922, an increase of 2.3%; among them, the mAP@0.5 index of airplane has increased by 1.1%, the mAP@0.5 index of oil tank has increased by 2%, the mAP@0.5 index of bridge has increased by 4.2%, and finally the mAP@0.5 index of ship has increased by 1.7%. However, due to the improvement of the C2f module, the number of model parameters has also increased from 3006428 to 3137821, which has the effect of increasing the inference speed by 0.3ms. Although the method of the present invention has a slight increase in the number of parameters, the increase is not large. The improvement of the overall detection effect of the model on the four types of targets can better make up for the defect that the current SAR image target detection accuracy is not high.

[0111] The comparative experimental results show that the proposed method has certain advantages in performance compared with the original YOLOv8. The method shows better detection effect in multi-category SAR image target detection and has certain practical value.

[0112] Although the specific implementation of the invention is described in detail in conjunction with the drawings, it should not be understood as limiting the scope of protection of this patent. Within the scope described in the claims, various modifications and variations that can be made by those skilled in the art without creative work still fall within the scope of protection of this patent.

Claims

1. A SAR image multi-target detection method based on YOLOv8, characterized in that: The following steps are involved: S1. Improve the YOLOv8 model; S2. Use the public SAR image dataset to train the improved YOLOv8 model to obtain a lightweight SAR image detection model; S3, using lightweight SAR image detection model to identify multiple targets in SAR images; S4. Perform performance evaluation on lightweight SAR image detection model based on multi-target detection results.

2. The SAR image multi-target detection method based on YOLOv8 improvement according to claim 1 is characterized in that: The S1 comprises the following sub-steps: S11. Add the global maximum pooling path and the global average pooling path to the SPPF module in the YOLOv8 model backbone network, and add the CA coordinate attention mechanism after the output of the SPPF_improve module; S12, use dynamic upsampling DySample to replace the upsample Upsample in the YOLOv8 model; S13, use the dilated residual module DWR to replace the Bottleneck module in C2f in the YOLOv8 model; S14. Change the CIoU bounding box loss function in the YOLOv8 model to the Inner-CIoU loss function.

3. The SAR image multi-target detection method based on YOLOv8 improvement according to claim 2 is characterized in that: In S14, the Inner-CIoU loss function is: L Inner-CoU =L CIoU +IoU-IoU inner Where, L Inner-CoU is the Inner-CIoU loss function; L CIoU is the original CIoU loss function; IoU is the intersection-over-union ratio between the original predicted box and the real box; IoU inner is the embedded IoU, which represents the intersection over union (IoU) of the predicted box and the real box after the scaling factor is introduced.

4. The SAR image multi-target detection method based on YOLOv8 improvement according to claim 3 is characterized in that: The original CIoU loss function L CIoU for: Where B is the area of ​​the prediction box; B gt is the area of ​​the real box; υ is a parameter to measure the consistency of aspect ratio; w gt is the width of the real frame; h gt is the height of the real box; w is the width of the predicted box; h is the height of the predicted box; α is the balance factor; ρ is the Euclidean distance between the predicted box and the center point of the real box; b is the coordinate of the center point of the predicted box; b gt is the coordinate of the center point of the real box; c is the diagonal length of the minimum circumscribed rectangle of the predicted box and the real box.

5. The SAR image multi-target detection method based on YOLOv8 improvement according to claim 4 is characterized in that: Based on the original CIoU loss function L CIoU , introduce scaling auxiliary bounding box ratio and calculate IoU inner for: In the formula, are the left, right, top and bottom boundaries of the ground-truth box respectively; They are the horizontal and vertical coordinates of the center point of the real frame respectively; according to And the four values ​​​​b of the prediction box l 、b r 、b t and b b , calculate the intersection inter of the scaled prediction box and the true box and the union union of the scaled prediction box and the true box: union=(w gt ×h gt )×ratio 2 +(w×h)×ratio 2 -inter Calculate IoU based on inter and union inner for: Where inter is the intersection of the scaled prediction box and the true box; union is the union of the scaled prediction box and the true box.

6. The SAR image multi-target detection method based on YOLOv8 improvement according to claim 1, characterized in that: The SAR image dataset in S2 is the public MSAR-1.0 dataset, which is randomly divided into a training set, a validation set, and a test set in a ratio of 7:1:

2.

7. The SAR image multi-target detection method based on improved YOLOv8 according to claim 6, characterized in that: In S4, the test set is used to evaluate the performance of the lightweight SAR image detection model: Where Precision is the accuracy rate; TP is the number of positive classes predicted as positive classes; FP is the number of negative classes predicted as positive classes; FN is the number of positive classes predicted as negative classes; Recall is the recall rate; AP is the precision rate; mAP is the average precision rate; P(R) is the relationship curve between precision and recall rate with horizontal and vertical coordinates; R is the recall rate; n is the number of categories of the detected targets; AP is the average precision rate; i is the average precision of the ith category.

Images