SAR image ship target detection method based on improved YOLOV8 network

Abstract

The invention discloses an SAR image ship target detection method based on an improved YOLOV8 network, and relates to the technical field of computer vision, and the method comprises the following steps: obtaining a disclosed SAR ship image data set, and carrying out the preprocessing of the data set; configuring a model training environment; an improved YOLOV8 network model is constructed, the improved YOLOV8 network model takes a YOLOv8 network as a basic network, convolution modules except a first convolution module and a second convolution module in the YOLOv8 network are replaced by DRFD modules, and an SPPF module in a backbone network of the YOLOv8 network is replaced by a DESSPC module; loading the constructed improved YOLOV8 network model into a configured training environment, initializing a network model parameter file, adjusting an input image size, setting a training round number and a batch-size, and then training and verifying the network model by using a pre-divided training set and a pre-divided verification set; and taking a to-be-detected image as input, and performing SAR image ship target detection through the trained and verified improved YOLOV8 network model. According to the method, the problems of low SAR image ship detection precision, low detection speed and the like under complex conditions are solved.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to a SAR image ship target detection method based on an improved YOLOV8 network. Background Art

[0002] Synthetic Aperture Radar (SAR) is an active remote sensing technology that can generate high-resolution radar images at any time and in all weather conditions by emitting pulsed electromagnetic waves and receiving reflected signals. Due to its all-weather, continuous and anti-interference detection capabilities, SAR imaging technology is widely used in the field of marine monitoring and plays a key role in maintaining national marine security and promoting civilian production activities.

[0003] With the rapid development and maturity of deep learning, the SAR image ship detection method based on deep learning has become a hot research topic. This method provides higher detection accuracy and speed and stronger generalization ability. Current target detection algorithms, such as Faster R-CNN, DETR, SSD and other algorithms, have achieved good results in the field of target detection. However, since SAR images contain a large amount of speckle noise, when the background is complex, these noises and the metal scattering of the ship are intertwined to cause the target appearance to deform, which makes it difficult for the deep network to extract representative target features. In addition, the complex background of ship images, the large scale difference between large and small ships, and the high density of ships in the docking scene make the detection algorithm prone to missed detection and false detection. In addition, in the scene of SAR ship image detection with strong real-time requirements, most algorithms have greatly increased the amount of calculation and model volume while improving the accuracy. The model is lightweight without improving the accuracy. The detection algorithm still has a lot of room for improvement in detection accuracy and detection speed. Summary of the invention

[0004] In view of the deficiencies in the prior art, the present invention provides a SAR image ship target detection method based on an improved YOLOV8 network to solve the problems of low SAR image ship detection accuracy and slow detection speed under complex conditions.

[0005] In order to achieve the above technical purpose, the technical solution adopted is: a SAR image ship target detection method based on an improved YOLOV8 network, comprising the following steps: Step 1: Obtain and preprocess the public SAR ship image dataset; Step 2: Configure the model training environment; Step 3, construct an improved YOLOV8 network model, wherein the improved YOLOV8 network model uses the YOLOv8 network as the basic network, replaces all convolution modules except the first and second in the YOLOv8 network with DRFD modules, and replaces the SPPF module in the backbone network of the YOLOv8 network with a DESSPC module; Step 4, load the constructed improved YOLOV8 network model into the configured training environment, initialize the network model parameter file, adjust the input image size, set the number of training rounds and batch-size, and then use the pre-divided training set and verification set to train and verify the network model; Step 5: Take the image to be detected as input and perform SAR image ship target detection through the trained and verified improved YOLOV8 network model.

[0006] The preprocessing method in step 1 of the present invention is: formatting the acquired data set to obtain a data set that meets the requirements of the improved YOLOV8 network model, and then randomly dividing the data set that meets the requirements into a training set and a validation set in a ratio of 7:3.

[0007] The DRFD module described in the present invention has three branch steps of D, C and M. The D branch uses grouped convolution and deep convolution to downsample the initial feature map and extract features, and then adopts GELU activation function and batch normalization to obtain the D branch feature map; the M branch performs the maximum pooling downsampling operation on the feature map after the grouped convolution operation in the D branch to obtain the M branch feature map; the C branch performs a slicing downsampling operation on the initial feature map to obtain the C branch feature map; finally, the feature maps of the three downsampling branches are adaptively fused to generate the final output feature map.

[0008] The adaptive fusion formula of the present invention is expressed as: fиsion=BN(Conv(Concat(x,y,z))) Where fиsion represents the feature map generated after adaptive fusion, BN represents batch normalization, Conv represents ordinary convolution with a kernel size of 1×1, Concat represents feature concatenation, and x, y, and z represent the feature maps output by the three downsampling branches.

[0009] The construction method of the DESSPC module described in the present invention is: based on the SPPF module, detail enhancement convolution is used to replace the standard convolution in the SPPF module, and three expansion convolution layers with shared parameters of different expansion rates are used to replace the maximum pooling layer in the SPPF module.

[0010] Step 3 of the present invention also includes adopting a small target feature fusion structure in the neck network.

[0011] The specific implementation method of the small target feature fusion structure described in the present invention is: upsampling the 80×80 feature map in the neck network of the YOLOV8 network to 160×160, and then splicing the upsampled 160×160 feature map with the 160×160 feature map output by C2f in the backbone network, and then generating a P2 feature map through the C2f module, by downsampling the P2 feature map to 80×80, and then splicing the downsampled 80×80 feature map with the 80×80 feature map output by the C2f module in the neck network, and then outputting the P3 feature map through the C2f module; by downsampling the P3 feature map to 40×40, and then splicing the downsampled 40×40 feature map with the 40×40 feature map output by the C2f module in the neck network, and then outputting the P4 feature map through the C2f module.

[0012] Step 3 of the present invention also includes replacing the detection head of the head network with a shared parameter detection head for detail enhancement.

[0013] The implementation method of the detail enhanced shared parameter detection head described in the present invention is to add a 1×1 convolution structure after each of the feature maps generated by the neck network to adjust the number of channels of the feature maps to be consistent, and then use two shared parameter detail enhancement convolutions to extract features from the feature maps, and then calculate Box.loss and Cls.loss respectively.

[0014] The present invention adjusts the input image size to 640×640, sets the number of training rounds to 300, and the batch-size to 8.

[0015] The present invention has the following characteristics and beneficial effects: (1) In order to solve the problem that the model adopts convolution downsampling and retains fine-grained details poorly, resulting in blurred or lost target features, the present invention introduces a DRFD robust feature downsampling module in the backbone network. The module uses three downsampling methods to extract local features, global features and original features respectively, and obtains the final optimized output feature map by fusing the three extracted feature maps, thereby improving the network model's ability to extract effective features and reducing the risk of key information loss during the model downsampling process, thereby improving the detection accuracy of the model.

[0016] (2) In order to cope with the challenges brought by the large differences in ship scales and the overlapping of docked ships, the present invention replaces the spatial pooling pyramid layer in the backbone network with a shared spatial pyramid convolution layer with DESSPC detail enhancement. This module effectively perceives the gradient differences in the image, thereby greatly improving the problem of false detection of ships caused by overlapping ships. In addition, this module dynamically adjusts the size of the receptive field when extracting features, effectively improving the problem of missed detection of ships due to large differences in ship scales.

[0017] (3) In order to cope with the scenario where the target of small-sized ships is high in the ship detection, the present invention adopts a small target feature fusion structure in the neck network, which generates a P2 layer feature map and removes the P5 detection layer, so that the model focuses on the small-sized detection targets in the image, thereby greatly improving the model's detection accuracy for small-sized targets.

[0018] (4) In order to reduce the computational complexity of the network model and enable the three detection heads to interact with each other, the present invention uses a new type of shared partial parameter detection head to replace the original detection head in the head network. This innovation can reduce the computational complexity of the network model while automatically adjusting its internal parameters according to the size of the input image and the changes in the feature map, thereby improving the detection capability of multi-scale input. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is the original YOLOv8 model structure diagram.

[0020] Figure 2 This is a model structure diagram of the present invention.

[0021] Figure 3 for Figure 2 DRFD module structure diagram.

[0022] Figure 4 for Figure 2 DESSPC module structure diagram.

[0023] Figure 5 for Figure 2 Structure diagram of the shared parameter detection head with medium detail enhancement.

[0024] Figure 6 This is the original model detection diagram.

[0025] Figure 7 This is a test diagram of an example of the present invention. DETAILED DESCRIPTION

[0026] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0027] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings.

[0028] On the contrary, the present invention covers any substitution, modification, equivalent method and scheme made on the essence and scope of the present invention as defined by the claims. Further, in order to make the public have a better understanding of the present invention, some specific details are described in detail in the detailed description of the present invention below. Those skilled in the art can fully understand the present invention without the description of these details.

[0029] The present invention provides a SAR image ship target detection method based on a DE-YOLO network, comprising the following steps: Step 1: Obtaining the dataset This embodiment uses the HRSID dataset released by the University of Electronic Science and Technology of China in January 2020. The dataset contains 5604 high-resolution SAR images and 16951 ship targets. The dataset images are captured by the Gaofen-3 satellite, Sentinel-1, and TerraSAR-X, covering different types of ocean scenes such as ports, nearshore and deep sea. It is the current mainstream SAR image ship dataset. The images in the dataset are randomly divided into training set and validation set in a ratio of 7:3, of which the training set contains 3923 images and the test set contains 1681 images.

[0030] Step 2: Build a training environment Create a dataset.yaml configuration file for the dataset, and write the training set storage path, test set storage path, and detection target type and name into the configuration file. Configure the deep learning environment as follows: cuda11.8, deep learning framework pytorch2.0.1, AMD Ryzen7 6800H CPU, 16G memory, NVIDIA GeForce RTX 3060 GPU, 6G video memory. Set the input image size to 640×640, the number of training rounds to 300, the batch size to 8, and keep the rest of the parameters as default.

[0031] Step 3: Build and improve the YOLOV8 network model, such as Figure 1 , Figure 2 As shown, the present invention mainly makes the following improvements. The improved YOLOv8 network model uses the YOLOv8 network as the basic network, and can improve four points, namely: (1) replacing the convolution modules except the first and second convolution modules in the YOLOv8 network with DRFD modules; (2) replacing the SPPF module in the backbone network with a DESSPC module; (3) adopting a small target feature fusion structure in the neck network; (4) replacing the detection head of the head network with a shared parameter detection head with detail enhancement.

[0032] (1) The present invention introduces the DRFD robust feature downsampling module to replace the last three downsampling convolution modules in the backbone network, as shown in the attached figure. Figure 3As shown in the figure, the DRFD module is divided into three downsampling branches according to different downsampling methods, which are denoted as D, C and M respectively; the D branch uses group convolution and depth convolution to downsample the initial feature map to extract features, reduce the computational complexity and fuse the global features of the feature map, and then uses GELU activation function and batch normalization to obtain the D branch feature map, which effectively accelerates the convergence of the network model and improves the stability of the model; the M branch performs the maximum pooling downsampling operation on the feature map after the group convolution operation in the D branch, extracts the significant features of the feature map, and obtains the M branch feature map; the C branch performs a slice downsampling operation on the initial feature map, and by cutting adjacent pixels, reduces the size of the feature map while retaining the original feature information, and obtains the C branch feature map; finally, the feature maps of the three downsampling branches are adaptively fused to generate the final output feature map.

[0033] The convolutional downsampling branch (D branch) first uses a grouped convolution with a stride of 1 and a kernel size of 3×3 to change the number of channels C of the input feature map to the number of output channels 2C specified in the model configuration file, and then uses a deep convolution with a stride of 2 and a kernel size of 3×3 for downsampling. Finally, the Gaussian Error Linear Unit (GELU) activation function and batch normalization are used to improve the training speed and stability of the network model. The convolutional downsampling operation extracts global feature information, and uses grouped convolution and deep convolution to effectively reduce the amount of model calculation. The maximum pooling downsampling branch (M branch) performs a 2×2 maximum pooling downsampling operation with a stride of 2 on the feature map after the grouped convolution operation in the convolutional downsampling branch. Maximum pooling downsampling can effectively extract local significant features and help the generalization ability of the model. The slice downsampling branch (C branch) uses the slicing method to downsample the input feature map. Specifically, by cutting adjacent pixels in the feature map, the feature map is divided into four feature maps with half the size of the original. Then the four feature maps are spliced ​​together using a splicing operation, which effectively reduces the size of the feature map and retains the original features. Then, a 1×1 convolution is used to adjust the number of channels of the spliced ​​slice feature map to 2C. Finally, as shown in the attached figure Figure 3 As shown, the feature maps of the three down-sampling branches are adaptively fused, and the fusion formula is shown in formula (1): fиsion=BN(Conv(Concat(x,y,z))) (1) In formula (1), fиsion represents the feature map generated after adaptive fusion, BN represents batch normalization, Conv represents ordinary convolution with a kernel size of 1×1, Concat represents feature concatenation, and x, y, and z represent the feature maps output by the three downsampling branches.

[0034] The DRFD module extracts and fuses the significant features, global features and original features in the image, achieving more efficient information retention and feature extraction, effectively capturing key information in the image, and thus improving the detection capability of the network model.

[0035] (2) The present invention proposes a novel detail-enhanced shared spatial pyramid convolutional layer (DESSPC) to replace the spatial pyramid pooling layer (SPPF module) in the backbone network, as shown in the attached figure. Figure 4 As shown in the figure, the DESSPC module uses detail enhancement convolution to replace the 1×1 convolution in the original spatial pyramid pooling layer, so that the model can not only extract basic features but also effectively extract gradient-level features, thereby enhancing the model's ability to capture details and edge information. Figure 4 As shown in the figure, the detail enhancement convolution contains five convolution layers deployed in parallel, namely: a regular convolution, a center difference convolution, an angle difference convolution, a horizontal difference convolution and a vertical difference convolution. The difference convolution first captures the difference between pixels in the image, and then uses the convolution kernel weight to perform the convolution operation to generate a feature map. Unlike regular convolution, the difference convolution not only considers the value of a single pixel when calculating the feature map, but also obtains deeper spatial information by calculating the difference between pixels. Compared with ordinary convolution, this method can enhance the representation and generalization capabilities of the model. In the design of detail enhancement convolution, ordinary convolution is used to obtain intensity level information, and difference convolution is used to obtain gradient level information in all directions. Finally, the features learned by the five convolutions are simply added to obtain the output of the detail enhancement convolution. By integrating the idea of ​​difference convolution, the detail enhancement convolution not only enhances the feature learning ability of the network, but also improves the model's ability to capture details in complex backgrounds. The specific calculation process of difference convolution is shown in formula (2): In formula (2), y(i,j) represents the pixel value of the convolution output, x(i,j) represents the pixel value of the input image at position (i,j), k represents the half-size of the convolution kernel, the convolution kernel size is (2k+1)×(2k+1), and w(m,n) represents the weight of the convolution kernel.

[0036] In order to further improve the model's detection capability for ship targets with large scale differences, the present invention uses three dilated convolutions with shared parameters of different dilation rates to replace the maximum pooling layer in the original spatial pyramid pooling layer. The maximum pooling operation is used to realize the model's extraction of significant features, but it will also lead to the loss of detailed features in the image, and the use of a fixed-size pooling window is not ideal when dealing with scenes with large changes in target scales such as ship detection. The dilated convolution inserts holes between the convolution kernels, thereby increasing the receptive field of the convolution operation without changing the size of the convolution kernel. By adjusting the size of the dilation factor, features of different scales can be extracted. Low dilation rates extract small target features, and high dilation rates extract large target features. Such a combination is conducive to extracting information of different sizes and different contexts in the image. The use of convolution parameter sharing also helps the model to realize the interaction of multi-scale information and improve the model's detection capability in complex scenes. The specific calculation process of the dilated convolution receptive field is shown in formula (3): R dilated =(k-1)×d+1 (3) In formula (3), R dilated represents the receptive field of the dilated convolution, k represents the convolution kernel size of the dilated convolution, and d represents the dilation factor of the dilated convolution.

[0037] (3) After a large number of convolution and pooling operations, the input image will cause some small information to be lost. The SAR image ship data contains many small target ships, which will greatly increase the probability of missed detection and false detection of ships. In order to solve the above problems, the present invention proposes a small target feature fusion structure in the neck network, as shown in the attached figure. Figure 1As shown in the figure. Based on the original YOLOv8 model, this structure first upsamples the 80×80 feature map output by the neck network to 160×160, then concatenates the upsampled 160×160 feature map with the 160×160 in the backbone network, and then generates a new output feature map P2 through the first C2f module on the right. In order to keep the model lightweight, the branch of concatenating the 20×20 feature map output by the backbone network with the 20×20 feature map of the neck network is removed, and the branch of generating the P5 feature map after downsampling, feature concatenation and C2f after the P4 feature map is removed. The P2 feature map is downsampled to 80×80, and then the downsampled 80×80 feature map is concatenated with the 80×80 feature map output by the first C2f module on the right in the neck network, and then the P3 feature map is output through the second C2f module on the right; the P3 feature map is downsampled to 40×40, and then the downsampled 40×40 feature map is concatenated with the 40×40 feature map output by the second C2f module on the right in the neck network, and then the P4 feature map is output through the third C2f module on the right. Finally, the head network detects the three layers of feature maps of P2, P3 and P4, which significantly improves the detection accuracy of the model for small targets.

[0038] (4) The present invention replaces the original detection head with a detail-enhanced shared parameter detection head in the head network. The new detection head can extract more detailed target features, realize information interaction between feature maps of different scales, improve model accuracy, and reduce the amount of model calculation. Specifically, as shown in Figure (5), the new detection head first uses three independent 1×1 ordinary convolutions to adjust the channels of the P2, P3 and P4 feature maps output by the neck network respectively. After the number of channels of the feature maps is adjusted to be consistent, two shared parameter detail enhancement convolutions are then used for feature extraction. Finally, Box.loss and Cls.loss are calculated respectively. During the detection process, the new detection head can dynamically adjust its internal parameters according to the size ratio of the three input feature maps, thereby improving the adaptability to inputs of different sizes and the detection accuracy.

[0039] Step 4: Use the optimized network model for training The HRSID dataset is divided into a training set and a validation set in a ratio of 7:3. The number of training rounds is set to 300, and the number of input images for a single training is set to 8. During the training process, the training process is observed in real time through Tensorboard. After the training is completed, the trained weights are saved.

[0040] Step 5: Use the pre-trained network model for target detection Specifically, the output feature maps of the first three C2f modules and the DESSPC module in the backbone network of the improved YOLOV8 network model are used as the input of the neck network, and the output of the neck network is used as the input of the head network. In order to prove the effectiveness of the technical solution of the present invention, the effects achieved by this invention are described in detail below in combination with the results of the ablation experiment. The experimental results are shown in Table 1, where Small represents the small target feature fusion structure, DESHead represents the shared detection head with detail enhancement, and √ indicates that the module is adopted. After the DRFD module is introduced, compared with the baseline model YOLOv8n, mAP 50 Improved by 0.6 percentage points, mAP 50-95 The mAP increased by 0.9 percentage points, the recall rate increased by 0.5 percentage points, and the amount of calculation increased slightly. After the introduction of the DRFD module and the DESSPC module, the mAP 50 Improved by 0.7 percentage points, mAP 50-95 The mAP increased by 1.4 percentage points, and the recall rate increased by 0.9 percentage points. The computational complexity remained the same as the baseline model. After introducing the DRFD module, DESSPC module and small target feature fusion structure, the mAP 50 Improved by 2 percentage points, mAP 50-95 The mAP of the detection head after introducing the DRFD module, DESSPC module, small target feature fusion structure and detail enhancement contribution parameter is improved by 4 percentage points, the recall rate is improved by 4.1 percentage points, and the computational complexity is improved by 41.4%. 50 Improved by 2.6 percentage points, mAP 50-95 The accuracy was improved by 4.2 percentage points, the recall rate was improved by 5.2 percentage points, and the computational complexity was reduced by 26.8% compared to the baseline model. Table 1 Ablation experiment

[0041] In order to demonstrate the effect of the invention, Figure 6 and Figure 7 Explanation: Compare the detection graph of the improved model with the original model. Figure 6 is the original model detection graph, Figure 7 To improve the model detection diagram. Figure 6 As shown in Figure 2, when the ship is approaching the shore, the original model misdetects the coast as a ship. Figure 7 The improved model has no false detections and can accurately detect each target. It can be seen that the improved model has better detection performance than the original model in the context of small target detection, ship docking, ship overlap, etc.

[0042] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions and variations of these embodiments including components are made without departing from the principles and spirit of the present invention, and still fall within the scope of protection of the present invention.

Claims

1. A SAR image ship target detection method based on an improved YOLOV8 network, characterized by: The following steps are involved: Step 1: Obtain and preprocess the public SAR ship image dataset; Step 2: Configure the model training environment; Step 3, construct an improved YOLOv8 network model, wherein the improved YOLOv8 network model uses the YOLOv8 network as the basic network, replaces all convolution modules except the first and second convolution modules in the YOLOv8 network with DRFD modules, and replaces the SPPF module in the backbone network of the YOLOv8 network with a DESSPC module; Step 4: Load the constructed improved YOLOV8 network model into the configured training environment, initialize the network model parameter file, adjust the input image size, set the number of training rounds and batch-size, and then use the pre-divided training set and verification set to train and verify the network model; Step 5: Take the image to be detected as input and perform SAR image ship target detection through the trained and verified improved YOLOV8 network model.

2. The SAR image ship target detection method based on the improved YOLOV8 network as claimed in claim 1, characterized in that: The preprocessing method in step 1 is: formatting the acquired data set to obtain a data set that meets the requirements of the improved YOLOV8 network model, and then randomly dividing the data set that meets the requirements into a training set and a validation set in a ratio of 7:

3.

3. The SAR image ship target detection method based on the improved YOLOV8 network as claimed in claim 1, characterized in that: The DRFD module has three branch steps of D, C and M. The D branch uses group convolution and depth convolution to downsample the initial feature map and extract the feature, and then uses GELU activation function and batch normalization to obtain the D branch feature map; the M branch performs the maximum pooling downsampling operation on the feature map after the group convolution operation in the D branch to obtain the M branch feature map; The C branch performs a slice down-sampling operation on the initial feature map to obtain the C branch feature map; finally, the feature maps of the three down-sampling branches are adaptively fused to generate the final output feature map.

4. The method for detecting ship targets in SAR images based on an improved YOLOV8 network as claimed in claim 3, characterized in that: The adaptive fusion formula is expressed as: Wherein, represents the feature map generated after adaptive fusion, represents batch normalization, represents ordinary convolution with a kernel size of 1×1, represents feature concatenation, and represents the feature map output by the three downsampling branches.

5. The SAR image ship target detection method based on the improved YOLOV8 network as claimed in claim 1, characterized in that: The construction method of the DESSPC module is as follows: on the basis of the SPPF module, the standard convolution in the SPPF module is replaced by the detail enhancement convolution, and the maximum pooling layer in the SPPF module is replaced by the dilated convolution layer with three shared parameters of different dilation rates.

6. The SAR image ship target detection method based on the improved YOLOV8 network as claimed in claim 1, characterized in that: The step 3 also includes adopting a small target feature fusion structure in the neck network.

7. The method for detecting ship targets in SAR images based on an improved YOLOV8 network as claimed in claim 6, characterized in that: The specific implementation method of the small target feature fusion structure is as follows: upsampling the 80×80 feature map in the neck network of the YOLOV8 network to 160×160, and then splicing the upsampled 160×160 feature map with the 160×160 feature map output by C2f in the backbone network, and then generating the P2 feature map through the C2f module, downsampling the P2 feature map to 80×80, and then splicing the downsampled 80×80 feature map with the 80×80 feature map output by the C2f module in the neck network, and then outputting the P3 feature map through the C2f module; downsampling the P3 feature map to 40×40, and then splicing the downsampled 40×40 feature map with the 40×40 feature map output by the C2f module in the neck network, and then outputting the P4 feature map through the C2f module.

8. The method for detecting ship targets in SAR images based on an improved YOLOV8 network as claimed in claim 1, characterized in that: The step 3 also includes replacing the detection head of the head network with a shared parameter detection head for detail enhancement.

9. The method for detecting ship targets in SAR images based on an improved YOLOV8 network as claimed in claim 8, characterized in that: The implementation method of the detail enhanced shared parameter detection head is as follows: a 1×1 convolution structure is added after each of the feature maps generated by the neck network to adjust the number of channels of the feature maps to be consistent, and then two shared parameter detail enhancement convolutions are used to extract features from the feature maps, and then Box.loss and Cls.loss are calculated respectively.

10. The method for detecting ship targets in SAR images based on an improved YOLOV8 network as claimed in claim 1, characterized in that: Adjust the input image size to 640×640, set the number of training rounds to 300, and the batch-size to 8.

Images