Ship wall surface treatment method

Abstract

The present invention provides a ship wall surface treatment method, which belongs to the field of ship wall surface treatment technology, and includes the following steps: S1, collecting ship wall surface defect images, constructing a ship wall surface defect data set and preprocessing, and obtaining a preprocessed data set; S2, inputting the preprocessed data set as an input image into an improved YOLOv8 model for feature extraction; S3, confirming the type of ship to be processed and the type of ship wall defect; S4, the processing center sends the ship type and ship wall defect information to a surface treatment robot, and the surface treatment robot identifies the surrounding environment information and formulates a surface treatment route; S5, processing the ship wall along the surface treatment route. The present invention adopts the above-mentioned ship wall surface treatment method, which can realize efficient and accurate processing of the ship wall, identify defects through image recognition technology, and combine with automatic robot path planning to realize intelligent and unmanned processing operations.

Description

Technical Field

[0001] The invention relates to the technical field of ship wall surface treatment, and in particular to a ship wall surface treatment method. Background Art

[0002] In the ship maintenance industry, ship wall surface treatment has always been a key link in maintaining the appearance of the ship, ensuring navigation safety and extending the service life of the ship. Traditionally, this task is mainly completed manually, but it faces many challenges, as follows:

[0003] (1) Manual processing is inefficient. For large ships, processing (cleaning) the entire ship wall requires a lot of time and manpower;

[0004] (2) Manual processing is difficult to ensure uniformity and accuracy, especially when dealing with complex or difficult-to-reach areas;

[0005] (3) Prolonged working at heights and exposure to hazardous processing substances pose a potential threat to workers’ health.

[0006] In recent years, with the rapid development of science and technology, especially the widespread application of image recognition, artificial intelligence and automation technology, the ship maintenance industry has brought revolutionary opportunities. These advanced technologies can not only improve work efficiency, but also achieve more refined operations while ensuring safety. Image recognition technology can accurately capture various defects on the surface of the ship wall and provide clear guidance for the surface treatment of the ship wall; artificial intelligence technology can intelligently analyze and process these defects and optimize the surface treatment strategy; and automated processing equipment can automatically complete the performance processing tasks according to the preset program, reducing manual intervention.

[0007] Therefore, the introduction of these advanced technologies in the field of ship maintenance can not only solve the problems existing in traditional surface treatment methods, but also promote the intelligent and automated upgrade of the entire industry and improve the overall maintenance level and efficiency. This provides broad space and prospects for the innovation and development of ship wall surface treatment technology. Summary of the invention

[0008] The purpose of the present invention is to provide a method for treating the surface of a ship wall, so as to achieve efficient and accurate treatment of the ship wall, identify defects through image recognition technology, and combine it with automatic robot path planning to realize intelligent and unmanned processing operations, thereby improving efficiency and safety.

[0009] To achieve the above object, the present invention provides a method for treating a ship wall surface, comprising the following steps:

[0010] S1, collecting ship wall surface defect images, constructing a ship wall surface defect data set and preprocessing it to obtain a preprocessed data set;

[0011] S2, input the preprocessed dataset as input image into the improved YOLOv8 model for feature extraction;

[0012] S3. Confirm the type of vessel to be processed and the type of ship wall defects;

[0013] S4. After confirming the ship type and the ship wall defect information, the processing center sends the ship type and the ship wall defect information to the surface treatment robot, and the surface treatment robot identifies the surrounding environment information and formulates a surface treatment route;

[0014] S5. Treat the ship wall along the surface treatment route.

[0015] Preferably, the ship wall surface defect image dataset includes surface scratches, surface rust, surface damage, paint peeling and surface attached barnacles.

[0016] Preferably, in step S1, the preprocessing includes the following sub-steps:

[0017] S11, randomly flipping, randomly cropping, randomly scaling and randomly translating the collected images of ship wall surface defects to obtain an expanded data set;

[0018] S12, graying the expanded data set to obtain a grayscale image;

[0019] S13. Processing the grayscale image through the image clarity enhancement model to obtain an image with clarity enhancement, which is the preprocessed data set.

[0020] Preferably, in step S13, the image definition enhancement module includes a scale variation structure, a residual dense block, a multi-scale feature extraction structure, a two-dimensional convolutional neural network and an edge detection structure;

[0021] Among them, the scale change structure includes an upsampling module and a downsampling module; the residual dense block includes three convolutional layers and an attention channel; the edge detection structure includes a horizontal edge detection network and a vertical edge detection network; the two-dimensional convolutional neural network includes three two-dimensional convolutional blocks, and the three two-dimensional convolutional blocks are composed of a Conv2d, BatchNorm2d, ReLU and MaxPool2d. The number of convolution kernels is 16, 32, and 64 respectively, and the size of the convolution kernel is , the core size of MaxPool2d is .

[0022] Preferably, in step S13, the image definition enhancement module works as follows:

[0023] The scale-varying structure first reduces the grayscale image and uses five residual dense blocks at different scales to extract multi-scale features to obtain the first feature map;

[0024] The scale-varying structure enlarges the size of the first feature map and uses five residual dense blocks at different scales to extract multi-scale features to obtain the second feature map.

[0025] Inputting the second feature map into the edge detection structure for edge detection processing to obtain an edge feature map of the grayscale image;

[0026] Input the edge feature map into the two-dimensional convolutional network for feature extraction to obtain the target feature map of the grayscale image;

[0027] The target feature map is fused with the grayscale image to obtain an image with improved clarity.

[0028] Preferably, in step S2, the preprocessed data set is input as an input image into the improved YOLOv8 network for feature extraction, including the following sub-steps:

[0029] S21. Use the backbone network of the improved YOLOv8 model to extract multi-scale feature maps from the input image;

[0030] S22, performing global average pooling and global maximum pooling on the multi-scale feature map;

[0031] The pooled result is transformed through the convolution layer to obtain the transformed features;

[0032] Use the Sigmoid activation function to convert the transformed feature map into a spatial attention weight matrix;

[0033] The calculation formula of the spatial attention weight matrix is:

[0034] ;

[0035] in, is the spatial attention weight matrix; It is a multi-scale feature map; is the Sigmoid activation function; It is a k×k convolution operation;

[0036] Multiply the spatial attention weight matrix with the multi-scale feature map to obtain the spatial weighted feature map;

[0037] S23, converting the spatial weighted feature map from the spatial domain to the frequency domain using discrete Fourier transform;

[0038] In the frequency domain, a multi-axis frequency domain attention mechanism is designed. The multi-axis frequency domain attention mechanism focuses on different frequency axes through multiple branches, and uses the fully connected layer to learn the frequency domain attention weights.

[0039] Multiply the frequency domain attention weight by the frequency domain feature to obtain the frequency domain weighted feature;

[0040] The frequency domain weighted features are converted back to the spatial domain through inverse discrete Fourier transform to obtain the feature map after frequency domain attention enhancement;

[0041] S24, fusing the spatial weighted feature map with the feature map enhanced by frequency domain attention;

[0042] S25. Use the detection head of the improved YOLOv8 model to perform target detection on the fused feature map and output the detection results.

[0043] Preferably, in step S21, the improved YOLOv8 model specifically refers to a backbone network of the YOLOv8 model enhanced by online reparameterization, and the improvement method is as follows:

[0044] S211, replacing the BN layer in the heavy parameter block with a linear scaling layer;

[0045] S212, fuse the convolutional layer and the linear scaling layer, and then add a GN layer after the residual connection;

[0046] S213. Convert and fuse all convolutional layers into a single 3×3 convolution to achieve equivalent conversion from multi-branch to single-branch.

[0047] Preferably, in step S3, confirming the type of ship to be processed and the type of ship wall defect includes the following sub-steps:

[0048] S31, collecting key information of different types of ships and constructing three-dimensional models, and forming a three-dimensional ship database with the three-dimensional models of different types of ships;

[0049] Key information about the vessel includes design drawings, size information, and material information;

[0050] S32. Before the vessel enters the workbench, a high-precision camera is used to take a picture of the vessel, and then the taken image is transmitted to a processing center, which recognizes the image and confirms the type of the vessel;

[0051] When the current vessel information is stored in the processing center, the vessel type information is sent to the surface processing robot;

[0052] When the processing center does not store the current ship information, key information of the current ship is collected, a three-dimensional model of the current ship is constructed, and the three-dimensional database of the ship is updated;

[0053] S33. When the ship arrives at the workbench, a high-precision camera installed on the side of the workbench is used to take a picture of the ship wall, and then the captured image is transmitted to the processing center. The processing center identifies and confirms the ship wall defects through the improved YOLOv8 model constructed in step S2.

[0054] Preferably, in step S4, the processing center sends the ship type and ship wall defect information to the surface treatment robot, and the surface treatment robot identifies the surrounding environment information and formulates a surface treatment route, including the following sub-steps:

[0055] S41, dividing the processing area into a number of unit areas according to the three-dimensional ship model established in step S3;

[0056] S42, performing path planning in each unit area;

[0057] S421. Acquire key information in each unit area, where the key information includes the starting point, the end point, and obstacle information;

[0058] S422, using the RTK system to obtain the position of the surface treatment robot;

[0059] S423, the surface treatment robot detects the surrounding environment, including the position and posture of obstacles, through a visual camera, and measures the distance to the obstacles through a laser ranging sensor;

[0060] S424, constructing a surrounding environment map based on the data obtained by the visual camera and the laser ranging sensor, and identifying obstacles, ships and safe areas;

[0061] S425, calculate the spatial distance from the starting point to the end point through the distance measurement algorithm, using The path search algorithm searches for a path from the starting point to the end point in the environment map;

[0062] Distance Metrics:

[0063] ;

[0064] in, Indicates distance; are the coordinates of the starting point; are the coordinates of the target point;

[0065] Path search algorithm:

[0066] ;

[0067] in, It is a heuristic function used to estimate the cost from node node to the target location; , and are the coordinates of the target location; , and is the coordinate of the node node;

[0068] S426, evaluating and sorting the searched paths based on the objective function, taking into account the length of the path, the distance to the obstacle, and the surface treatment tool head data;

[0069] S427. When the evaluation finds a better path, perform path optimization.

[0070] Preferably, in step S426, the objective function is as follows:

[0071] ;

[0072] in, is the objective function for evaluating the comprehensive performance of the path; , , is the weight coefficient; represents the path length; is a measure related to obstacle avoidance; Indicates the distance between the surface treatment tool head and the location to be treated.

[0073] Therefore, the present invention adopts the above-mentioned ship wall surface treatment method, and the beneficial technical effects are as follows:

[0074] (1) The pretreatment of the present invention has multiple effects.

[0075] First, because the image is grayed and sharpened in the preprocessing stage, the interference of color information is removed, and the contrast and edge features of the image are enhanced, which makes the subsequent image recognition model more focused on the geometric shape and structural features of the image. This processing method helps to enhance the robustness and stability of the model, so that it can maintain good performance under different lighting conditions, shooting angles and noise environments.

[0076] Second, the original image dataset has been significantly expanded through data augmentation techniques (such as random flipping, cropping, scaling, and translation), which not only increases the number of samples for training the model, but also improves the diversity and representativeness of the samples. This increase in diversity helps the model learn richer features, thereby improving its generalization ability to new images and reducing overfitting.

[0077] Third, the image clarity enhancement module effectively improves the clarity of the image through the comprehensive application of multi-scale feature extraction, residual dense blocks, two-dimensional convolutional neural networks and edge detection technologies. A clear image can more accurately reflect the true situation of the ship wall surface, reduce recognition errors caused by factors such as image blur and noise, and thus improve the precision and accuracy of defect recognition.

[0078] Fourth, the automated pre-processing and clarity enhancement process reduces the need for manual intervention and improves the automation of the entire ship wall treatment operation. This not only reduces labor costs, but also improves processing efficiency, allowing surface treatment operations to be completed faster and more accurately.

[0079] (2) The present invention adopts an improved YOLOv8 model for image recognition. With its efficient feature extraction capability and fast target detection speed, the model can quickly and accurately identify various defects from the ship wall image, thereby ensuring that the processing operations can be directly carried out on these defects, greatly improving the pertinence and efficiency of the processing;

[0080] At the same time, through the collected environmental information, the surface treatment robot can build a three-dimensional environmental model in real time, accurately plan the treatment route, avoid invalid movement, and further improve the efficiency of the treatment operation;

[0081] (3) The path planning of the present invention has multiple effects. First, with the help of high-precision sensors and RTK systems, the robot can perceive and build a three-dimensional model of the surrounding environment in real time and accurately, providing a solid foundation for accurate path planning. The path search algorithm and comprehensive performance evaluation mechanism enable the robot to quickly plan the optimal path in a complex environment, and comprehensively consider the path length, obstacle avoidance effect and processing accuracy to ensure efficient execution of the task. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] Figure 1 This is a flow chart of a ship wall surface treatment method of the present invention;

[0083] Figure 2 It is a diagram of the online reparameterization process;

[0084] Figure 3 Flowchart for path planning in a unit area. DETAILED DESCRIPTION

[0085] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0086] Embodiment 1

[0087] like Figure 1FIG. 1 is a flow chart of a ship wall surface treatment method according to the present invention, comprising the following steps:

[0088] S1. Collecting images of ship wall surface defects, constructing a ship wall surface defect dataset (the ship wall surface defect image dataset includes surface scratches, surface rust, surface damage, paint peeling and surface attached barnacles) and preprocessing to obtain a preprocessed dataset;

[0089] Preprocessing includes the following sub-steps:

[0090] S11, randomly flipping, randomly cropping, randomly scaling and randomly translating the collected images of ship wall surface defects to obtain an expanded data set;

[0091] S12, graying the expanded data set to obtain a grayscale image;

[0092] S13. Processing the grayscale image through the image clarity enhancement model to obtain an image with clarity enhancement, which is the preprocessed data set.

[0093] The image clarity enhancement module includes a scale variation structure, a residual dense block, a multi-scale feature extraction structure, a two-dimensional convolutional neural network, and an edge detection structure;

[0094] Among them, the scale change structure includes an upsampling module and a downsampling module; the residual dense block includes three convolutional layers and an attention channel; the edge detection structure includes a horizontal edge detection network and a vertical edge detection network; the two-dimensional convolutional neural network includes three two-dimensional convolutional blocks, and the three two-dimensional convolutional blocks are composed of a Conv2d, BatchNorm2d, ReLU and MaxPool2d. The number of convolution kernels is 16, 32, and 64 respectively, and the size of the convolution kernel is , the core size of MaxPool2d is .

[0095] The image clarity enhancement module works as follows:

[0096] The scale-varying structure first reduces the grayscale image and uses five residual dense blocks at different scales to extract multi-scale features to obtain the first feature map;

[0097] The scale-varying structure enlarges the size of the first feature map and uses five residual dense blocks at different scales to extract multi-scale features to obtain the second feature map.

[0098] Inputting the second feature map into the edge detection structure for edge detection processing to obtain an edge feature map of the grayscale image;

[0099] Input the edge feature map into the two-dimensional convolutional network for feature extraction to obtain the target feature map of the grayscale image;

[0100] The target feature map is fused with the grayscale image to obtain an image with improved clarity.

[0101] S2. Input the preprocessed dataset as input image into the improved YOLOv8 model for feature extraction, including the following sub-steps:

[0102] S21. Use the backbone network of the improved YOLOv8 model to extract multi-scale feature maps from the input image;

[0103] like Figure 2 As shown in the figure, the improved YOLOv8 model specifically refers to the backbone network of the YOLOv8 model enhanced by online reparameterization. The improvement method is as follows:

[0104] S211, replacing the BN layer in the heavy parameter block with a linear scaling layer;

[0105] S212, fuse the convolutional layer and the linear scaling layer, and then add a GN layer after the residual connection;

[0106] S213. Convert and fuse all convolutional layers into a single 3×3 convolution to achieve equivalent conversion from multi-branch to single-branch.

[0107] S22, performing global average pooling and global maximum pooling on the multi-scale feature map;

[0108] The pooled result is transformed through the convolution layer to obtain the transformed features;

[0109] Use the Sigmoid activation function to convert the transformed feature map into a spatial attention weight matrix;

[0110] The calculation formula of the spatial attention weight matrix is:

[0111] ;

[0112] in, is the spatial attention weight matrix; It is a multi-scale feature map; is the Sigmoid activation function; It is a k×k convolution operation;

[0113] Multiply the spatial attention weight matrix with the multi-scale feature map to obtain the spatial weighted feature map;

[0114] S23, converting the spatial weighted feature map from the spatial domain to the frequency domain using discrete Fourier transform;

[0115] In the frequency domain, a multi-axis frequency domain attention mechanism is designed. The multi-axis frequency domain attention mechanism focuses on different frequency axes through multiple branches, and uses the fully connected layer to learn the frequency domain attention weights.

[0116] Multiply the frequency domain attention weight by the frequency domain feature to obtain the frequency domain weighted feature;

[0117] The frequency domain weighted features are converted back to the spatial domain through inverse discrete Fourier transform to obtain the feature map after frequency domain attention enhancement;

[0118] S24, fusing the spatial weighted feature map with the feature map enhanced by frequency domain attention;

[0119] S25. Use the detection head of the improved YOLOv8 model to perform target detection on the fused feature map and output the detection results.

[0120] S3. Confirm the type of vessel to be processed and the type of ship wall defects, including the following sub-steps:

[0121] S31, collecting key information of different types of ships and constructing three-dimensional models, and forming a three-dimensional ship database with the three-dimensional models of different types of ships;

[0122] Key information about the vessel includes design drawings, size information, and material information;

[0123] S32. Before the vessel enters the workbench, a high-precision camera is used to take a picture of the vessel, and then the taken image is transmitted to a processing center, which recognizes the image and confirms the type of the vessel;

[0124] When the current vessel information is stored in the processing center, the vessel type information is sent to the surface processing robot;

[0125] When the processing center does not store the current ship information, key information of the current ship is collected, a three-dimensional model of the current ship is constructed, and the three-dimensional database of the ship is updated;

[0126] S33. When the ship arrives at the workbench, a high-precision camera installed on the side of the workbench is used to take a picture of the ship wall, and then the captured image is transmitted to the processing center. The processing center identifies and confirms the ship wall defects through the improved YOLOv8 model constructed in step S2.

[0127] S4. After confirming the ship type and the ship wall defect information, the processing center sends the ship type and the ship wall defect information to the surface treatment robot. The surface treatment robot identifies the surrounding environment information and formulates a surface treatment route, including the following sub-steps:

[0128] S41, dividing the processing area into a number of unit areas according to the three-dimensional ship model established in step S3;

[0129] like Figure 3 As shown, S42, path planning is performed in each unit area;

[0130] S421. Acquire key information in each unit area, where the key information includes the starting point, the end point, and obstacle information;

[0131] S422, using the RTK system to obtain the position of the surface treatment robot;

[0132] S423, the surface treatment robot detects the surrounding environment, including the position and posture of obstacles, through a visual camera, and measures the distance to the obstacles through a laser ranging sensor;

[0133] S424, constructing a surrounding environment map based on the data obtained by the visual camera and the laser ranging sensor, and identifying obstacles, ships and safe areas;

[0134] S425, calculate the spatial distance from the starting point to the end point through the distance measurement algorithm, using The path search algorithm searches for a path from the starting point to the end point in the environment map;

[0135] Distance Metrics:

[0136] ;

[0137] in, Indicates distance; are the coordinates of the starting point; are the coordinates of the target point;

[0138] Path search algorithm:

[0139] ;

[0140] in, It is a heuristic function used to estimate the cost from node node to the target location; , and are the coordinates of the target location; , and is the coordinate of the node node;

[0141] S426, evaluating and sorting the searched paths based on the objective function, taking into account the length of the path, the distance to the obstacle, and the surface treatment tool head data;

[0142] ;

[0143] in, is the objective function for evaluating the comprehensive performance of the path; , , is the weight coefficient; represents the path length; is a measure related to obstacle avoidance; Indicates the distance between the surface treatment tool head and the position to be treated;

[0144] S427. When the evaluation finds a better path, the path is optimized (the path can be smoothed by interpolation method).

[0145] S5. Treat the ship wall along the surface treatment route.

[0146] The surface treatment robot (cleaning robot) mentioned in this embodiment includes a main body, on which a visual camera, a laser ranging sensor, a pressure sensor, a robotic arm, a surface treatment tool head, etc. are provided. The surface treatment robot and other components not described in detail are all prior arts.

[0147] Therefore, the present invention adopts the above-mentioned ship wall surface treatment method, identifies defects through image recognition technology, and combines it with automated robot path planning to realize intelligent and unmanned processing operations, thereby improving efficiency and safety.

[0148] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for treating a ship wall surface, characterized in that: The following steps are involved: S1, collecting ship wall surface defect images, constructing a ship wall surface defect data set and preprocessing it to obtain a preprocessed data set; S2, input the preprocessed dataset as input image into the improved YOLOv8 model for feature extraction; S3. Confirm the type of vessel to be processed and the type of ship wall defects; S4. After confirming the ship type and the ship wall defect information, the processing center sends the ship type and the ship wall defect information to the surface treatment robot, and the surface treatment robot identifies the surrounding environment information and formulates a surface treatment route; S5, treating the ship wall along the surface treatment route; In step S2, the preprocessed dataset is input as an input image into the improved YOLOv8 network for feature extraction, which includes the following sub-steps: S21. Use the backbone network of the improved YOLOv8 model to extract multi-scale feature maps from the input image; S22, performing global average pooling and global maximum pooling on the multi-scale feature map; The pooled result is transformed through the convolution layer to obtain the transformed features; Use the Sigmoid activation function to convert the transformed feature map into a spatial attention weight matrix; The calculation formula of the spatial attention weight matrix is: ； in, is the spatial attention weight matrix; It is a multi-scale feature map; is the Sigmoid activation function; for Convolution operation; Multiply the spatial attention weight matrix with the multi-scale feature map to obtain the spatial weighted feature map; S23, converting the spatial weighted feature map from the spatial domain to the frequency domain using discrete Fourier transform; In the frequency domain, a multi-axis frequency domain attention mechanism is designed. The multi-axis frequency domain attention mechanism focuses on different frequency axes through multiple branches, and uses the fully connected layer to learn the frequency domain attention weights. Multiply the frequency domain attention weight by the frequency domain feature to obtain the frequency domain weighted feature; The frequency domain weighted features are converted back to the spatial domain through inverse discrete Fourier transform to obtain the feature map after frequency domain attention enhancement; S24, fusing the spatial weighted feature map with the feature map enhanced by frequency domain attention; S25. Use the detection head of the improved YOLOv8 model to perform target detection on the fused feature map and output the detection result. The improved YOLOv8 model specifically refers to the backbone network of the YOLOv8 model enhanced by online reparameterization. The improvement method is as follows: S211, replacing the BN layer in the heavy parameter block with a linear scaling layer; S212, fuse the convolutional layer and the linear scaling layer, and then add a GN layer after the residual connection; S213, transform and merge all convolutional layers into a single 3×3 convolution to achieve equivalent conversion from multi-branch to single-branch; In step S3, the type of ship to be processed and the type of ship wall defect are confirmed, including the following sub-steps: S31, collecting key information of different types of ships and constructing three-dimensional models, and forming a three-dimensional ship database with the three-dimensional models of different types of ships; Key information about the vessel includes design drawings, size information, and material information; S32. Before the vessel enters the workbench, a high-precision camera is used to take a picture of the vessel, and then the taken image is transmitted to a processing center, which recognizes the image and confirms the type of the vessel; When the current vessel information is stored in the processing center, the vessel type information is sent to the surface processing robot; When the processing center does not store the current ship information, key information of the current ship is collected, a three-dimensional model of the current ship is constructed, and the three-dimensional database of the ship is updated; S33. When the ship arrives at the workbench, a high-precision camera installed on the side of the workbench is used to take a picture of the ship wall, and then the captured image is transmitted to the processing center. The processing center identifies and confirms the ship wall defects through the improved YOLOv8 model constructed in step S2.

2. A ship wall surface treatment method according to claim 1, characterized in that: The ship wall surface defect image dataset includes surface scratches, surface rust, surface damage, paint peeling, and surface attached barnacles.

3. A ship wall surface treatment method according to claim 2, characterized in that: In step S1, the preprocessing includes the following sub-steps: S11, randomly flipping, randomly cropping, randomly scaling and randomly translating the collected images of ship wall surface defects to obtain an expanded data set; S12, graying the expanded data set to obtain a grayscale image; S13. Processing the grayscale image through the image clarity enhancement model to obtain an image with clarity enhancement, which is the preprocessed data set.

4. A ship wall surface treatment method according to claim 3, characterized in that: In step S13, the image definition enhancement module includes a scale variation structure, a residual dense block, a multi-scale feature extraction structure, a two-dimensional convolutional neural network and an edge detection structure; Among them, the scale change structure includes an upsampling module and a downsampling module; the residual dense block includes three convolutional layers and an attention channel; the edge detection structure includes a horizontal edge detection network and a vertical edge detection network; the two-dimensional convolutional neural network includes three two-dimensional convolutional blocks, and the three two-dimensional convolutional blocks are composed of a Conv2d, BatchNorm2d, ReLU and MaxPool2d. The number of convolution kernels is 16, 32, and 64 respectively, and the size of the convolution kernel is , the core size of MaxPool2d is .

5. A ship wall surface treatment method according to claim 4, characterized in that: In step S13, the image definition enhancement module works as follows: The scale-varying structure first reduces the grayscale image and uses five residual dense blocks at different scales to extract multi-scale features to obtain the first feature map; The scale-varying structure enlarges the size of the first feature map and uses five residual dense blocks at different scales to extract multi-scale features to obtain the second feature map. Inputting the second feature map into the edge detection structure for edge detection processing to obtain an edge feature map of the grayscale image; Input the edge feature map into the two-dimensional convolutional network for feature extraction to obtain the target feature map of the grayscale image; The target feature map is fused with the grayscale image to obtain an image with improved clarity.

6. A ship wall surface treatment method according to claim 5, characterized in that: In step S4, the processing center sends the ship type and ship wall defect information to the surface treatment robot, and the surface treatment robot identifies the surrounding environment information and formulates a surface treatment route, including the following sub-steps: S41, dividing the processing area into a number of unit areas according to the three-dimensional ship model established in step S3; S42, performing path planning in each unit area; S421. Acquire key information in each unit area, where the key information includes the starting point, the end point, and obstacle information; S422, using the RTK system to obtain the position of the surface treatment robot; S423, the surface treatment robot detects the surrounding environment, including the position and posture of obstacles, through a visual camera, and measures the distance to the obstacles through a laser ranging sensor; S424, constructing a surrounding environment map based on the data obtained by the visual camera and the laser ranging sensor, and identifying obstacles, ships and safe areas; S425, calculate the spatial distance from the starting point to the end point through the distance measurement algorithm, using The path search algorithm searches for a path from the starting point to the end point in the environment map; Distance Metrics: ； in, Indicates distance; are the coordinates of the starting point; are the coordinates of the target point; Path search algorithm: ； in, It is a heuristic function used to estimate the cost from node node to the target location; , and are the coordinates of the target location; , and is the coordinate of the node node; S426, evaluating and sorting the searched paths based on the objective function, taking into account the length of the path, the distance to the obstacle, and the surface treatment tool head data; S427. When the evaluation finds a better path, perform path optimization.

7. A ship wall surface treatment method according to claim 6, characterized in that: In step S426, the objective function is as follows: ； in, is the objective function for evaluating the comprehensive performance of the path; , , is the weight coefficient; represents the path length; is a measure related to obstacle avoidance; Indicates the distance between the surface treatment tool head and the location to be treated.

Images