Ship anti-corrosion spraying quality detection method and system based on machine vision

Abstract

The invention belongs to the field of image processing, and provides a ship anticorrosion spraying quality detection method and system based on machine vision, and the method comprises the steps: obtaining a 360-degree panoramic image of a to-be-detected space of a ship body; performing structural semantic recognition on the panoramic image, and dividing the panoramic image into a plurality of regions with structural semantic tags based on a semantic segmentation network; establishing a structure-spraying process mapping rule, and forming a mapping relation between the structure type and the corresponding spraying type; local image acquisition is carried out on the ship body spraying completion area; carrying out spatial position reverse positioning on the local image, and obtaining a corresponding structure semantic tag; according to the structure semantic label, determining a spraying type and a spraying quality detection standard corresponding to the local image; and carrying out spraying quality detection on the local image, carrying out quality analysis on a spraying area in the local image by adopting a defect detection model based on image recognition, and outputting a quality evaluation result.

Description

Technical Field

[0001] This invention belongs to the field of image processing, and specifically relates to a method and system for inspecting the quality of anti-corrosion spraying on ships based on machine vision. Background Technology

[0002] Marine anti-corrosion spraying refers to applying a coating system with anti-corrosion functions to the surface of the ship's steel structure to resist corrosion damage caused by factors such as seawater, humidity, ultraviolet rays, and marine organisms, thereby extending the ship's service life and ensuring structural safety. The coating system typically consists of a primer, intermediate coat, and topcoat. Different types of coating materials and thicknesses are used for different parts depending on their function and environmental differences, such as zinc-rich primers, thick-film epoxy paints, or antifouling coatings.

[0003] Due to the complex and diverse structure of ship hulls, the wide distribution of spraying areas, and significant environmental differences, the requirements for anti-corrosion spraying vary from place to place. For example, the outer plating above the waterline is exposed to wind, rain, and sun, requiring a topcoat that is resistant to ultraviolet rays and has strong colorfastness; areas below the waterline need to resist long-term seawater immersion and microbial adhesion, necessitating the use of a coating system with antifouling capabilities; and internal structures such as ballast tanks and cargo holds have strict requirements for adhesion, abrasion resistance, and impermeability. Therefore, spraying strategies need to be customized based on the specific structural location.

[0004] Traditional manual sampling methods are insufficient for large-area, high-frequency coating quality assessment. They rely heavily on experience, are subjective, inefficient, and prone to missed or incorrect assessments. Machine vision-based methods for inspecting marine anti-corrosion coating quality enable automated, digital, and objective inspection of the coated area. Through image recognition and intelligent analysis, they determine the presence of defects such as missed areas, blistering, runs, and color differences, significantly improving inspection coverage and accuracy. This is a crucial supporting technology for intelligent shipbuilding and automated shipyard management.

[0005] However, this method still faces significant challenges in engineering implementation. On the one hand, while using large-scale images (such as panoramic images or large-format photographs) can cover a wide area, resolution limitations make it difficult to clearly identify detailed features such as coating defects. On the other hand, while using local images can help with detailed quality inspection, the complex structure of the hull makes it difficult to extract semantic structure, and isolated images cannot be automatically identified in terms of their structural region and coating type, requiring manual annotation or auxiliary information, resulting in a large workload and low efficiency. Summary of the Invention

[0006] To address the problems in the existing technology, this invention provides a machine vision-based method for inspecting the quality of ship anti-corrosion spraying, comprising the following steps:

[0007] S1. Obtain a 360-degree panoramic image of the space to be inspected on the hull. Take panoramic photos of the external and / or internal structure of the ship using a panoramic imaging device to form panoramic image data unfolded on a spherical surface.

[0008] S2. Perform structural semantic recognition on the panoramic image, and divide the panoramic image into multiple regions with structural semantic labels based on a semantic segmentation network. Each semantic label corresponds to a specific structural part of the ship's hull.

[0009] S3. Based on the structural semantic tags, establish structure-spraying process mapping rules to form a mapping relationship between structure type and corresponding spraying type;

[0010] S4. Acquire local images by using handheld devices, mobile terminals, or inspection robots to collect local images of the area where the ship hull has been painted.

[0011] S5. Perform spatial location inversion on the local image, match the feature information of the local image with the panoramic image, determine the mapping position of the local image in the panoramic image, and obtain the corresponding structural semantic label.

[0012] S6. Based on the structural semantic tags, determine the spraying type and spraying quality inspection standard corresponding to the local image;

[0013] S7. Perform spraying quality inspection on the local image, use an image recognition-based defect detection model to analyze the quality of the sprayed area in the local image, and output the quality assessment result.

[0014] In another aspect, the present invention provides a machine vision-based ship anti-corrosion spraying quality inspection system, comprising the following modules:

[0015] The panoramic image acquisition module is used to acquire 360-degree panoramic images of the space to be inspected on the ship's hull. It takes panoramic pictures of the ship's external and / or internal structure through panoramic imaging equipment to form panoramic image data unfolded on a spherical surface.

[0016] The structural semantic recognition module is used to perform structural semantic recognition on the panoramic image. Based on the semantic segmentation network, the panoramic image is divided into multiple regions with structural semantic labels, and each semantic label corresponds to a specific structural part of the ship's hull.

[0017] The spraying rule matching module is used to establish structure-spraying process mapping rules based on the structure semantic tags, forming a mapping relationship between structure type and corresponding spraying type;

[0018] The local image acquisition module is used to acquire local images of the painted area of ​​the ship hull using a handheld device, mobile terminal, or inspection robot.

[0019] The spatial location anti-localization module is used to match the feature information of the local image with the panoramic image, determine the mapping position of the local image in the panoramic image, and obtain the corresponding structural semantic label.

[0020] The spraying standard extraction module is used to determine the spraying type and spraying quality inspection standard corresponding to the local image based on the structural semantic tags.

[0021] The quality inspection module is used to inspect the spraying quality of the local image. It uses an image recognition-based defect detection model to analyze the quality of the sprayed area in the local image and outputs the quality assessment results.

[0022] Furthermore, the local image acquisition module includes a pose perception module that works synchronously with it. The pose perception module is one or a combination of an IMU (Inertial Measurement Unit), a QR code recognition system, or a visual-inertial positioning module.

[0023] The pose perception module is used to simultaneously record acceleration and angular velocity data when acquiring local images, and optimize the three-dimensional trajectory and attitude information of the computing device through Kalman filtering or sliding window as the initial value for spatial anti-positioning.

[0024] This invention integrates information from panoramic and local images to construct a mapping model between structural semantics and spraying strategies, achieving structure-location-driven spraying inspection process control. Symmetrical loss improves the accuracy of structure recognition, and panoramic structure recognition and spatial reverse positioning mechanisms automatically infer the structural location and spraying type of a local image, thereby loading targeted detection parameters and improving the engineering adaptability of the detection results.

[0025] This invention employs a deep learning-based defect identification model to determine coating defects. By combining structural semantic tags with a standard library of dynamically adapted quality assessment standards, it can accurately determine the location, type, and severity of defects in the coating area, thereby improving the automation level of detection and reducing reliance on manual labor.

[0026] The technical solution described in this invention supports rapid on-site deployment and edge device operation, possesses excellent real-time performance, scalability, and versatility, and is suitable for the testing needs of different ship types and different coating systems, thus having significant industrial promotion value. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is the main flowchart of the present invention;

[0029] Figure 2 This is a flowchart of the semantic segmentation process of this invention;

[0030] Figure 3 This is a flowchart of the reverse positioning process of the present invention. Detailed Implementation

[0031] The invention will now be described in preferred form with reference to the accompanying drawings and specific embodiments.

[0032] In order to solve the technical problems in the background art, such as Figure 1 As shown, this invention provides a machine vision-based method for inspecting the quality of anti-corrosion spraying on ships, belonging to the field of image processing, in order to improve the accuracy and efficiency of quality inspection.

[0033] The method includes the following steps:

[0034] S1. Obtain a 360-degree panoramic image of the space to be inspected on the hull. Take panoramic photos of the external and / or internal structure of the ship using a panoramic imaging device to form panoramic image data unfolded on a spherical surface.

[0035] S2. Perform structural semantic recognition on the panoramic image, and divide the panoramic image into multiple regions with structural semantic labels based on a semantic segmentation network. Each semantic label corresponds to a specific structural part of the ship's hull.

[0036] S3. Based on the structural semantic tags, establish structure-spraying process mapping rules to form a mapping relationship between structure type and corresponding spraying type;

[0037] S4. Acquire local images by using handheld devices, mobile terminals, or inspection robots to collect local images of the area where the ship hull has been painted.

[0038] S5. Perform spatial location inversion on the local image, match the feature information of the local image with the panoramic image, determine the mapping position of the local image in the panoramic image, and obtain the corresponding structural semantic label.

[0039] S6. Based on the structural semantic tags, determine the spraying type and spraying quality inspection standard corresponding to the local image;

[0040] S7. Perform spraying quality inspection on the local image, use an image recognition-based defect detection model to analyze the quality of the sprayed area in the local image, and output the quality assessment result.

[0041] The following is a further detailed explanation of each of the above steps.

[0042] S1. Obtain a 360-degree panoramic image of the space to be inspected on the hull. Use panoramic imaging equipment to take panoramic pictures of the external and / or internal structure of the ship to form panoramic image data unfolded on a spherical surface.

[0043] To achieve automated, structure-related detection of ship anti-corrosion coating quality, it is essential to collect comprehensive structural information about the space in which the hull is located before the entire inspection process begins, establishing a semantic mapping relationship between images and structures. Due to the complex hull structure, including extensive curved surfaces, deep interior compartments, and blind spots, images from a single viewpoint cannot fully cover the target inspection space and cannot accurately determine the structural location attributes of subsequent local images. Therefore, it is necessary to first acquire a panoramic image of the space to be inspected and perform structural semantic segmentation on this image, providing prior structural support for the structural localization and coating type determination in subsequent local images.

[0044] In this step, a panoramic image refers to image data that expresses full-view information of a space in a spherical or cylindrical unfolded manner, capable of covering image information within a horizontal 360-degree and vertical 180-degree range from the current shooting position. A panoramic imaging device is a device that can complete multi-directional synchronous imaging at a single point in time or within a short period of time, and output a panoramic image through stitching or reflection unfolding. Common devices include spherical cameras, structured light camera arrays, and fisheye camera groups.

[0045] The specific implementation process of step S1 is as follows: First, a panoramic imaging device is deployed in the area to be detected to collect images of the ship's external structure or internal compartments in a full-scene manner. In specific implementations, optional solutions include: using a spherical multi-lens camera (e.g., an omnidirectional camera device composed of six wide-angle cameras) to simultaneously acquire images from all directions at the same location; or using a rotatable high-resolution single lens with a mechanical turntable to capture images sequentially from multiple angles, and then generating a panoramic image through an image stitching algorithm. The stitching algorithm can be the SIFT feature point matching algorithm, the RANSAC model registration algorithm, or an image reprojection method based on spherical projection. Preferably, to improve the structural consistency and illumination balance of the image acquisition, the device can be equipped with an automatic exposure control module and an HDR imaging module to adapt to the low-light environment inside the ship and the strong reflective environment outside.

[0046] After image acquisition, the original image is geometrically unfolded using a spherical projection model to generate spherically unfolded image data. The principle of spherical unfolding is to map the spherical spatial coordinates (θ, φ) to two-dimensional image coordinates (x, y), and the mapping relationship is as follows:

[0047] x = W × (θ + π) ÷ 2π

[0048] y = H × (π ÷ 2 - φ) ÷ π

[0049] Where θ represents the horizontal angle (in radians), φ represents the vertical elevation angle (in radians), W is the horizontal resolution of the image, H is the vertical resolution of the image, and x and y are the coordinates of the corresponding pixel positions in the projected image.

[0050] In a preferred implementation, to enhance the stability of subsequent structural semantic recognition, the panoramic image data will be spatially registered according to preset hull design reference points (such as hatch numbers, structure numbers, or pre-embedded positioning tags) so that spatial back-projection of local images and determination of structural positions can be achieved subsequently.

[0051] In the example, for the corrosion inspection of the interior of a 200,000-ton bulk carrier, a spherical panoramic camera Insta360Pro2 was installed in the center of each ballast tank and captured images from four angles simultaneously. The captured images were then fused into a panoramic unfolded image with a resolution of 7680×3840 through a deep learning image stitching engine, which in turn formed the basic image data for structural semantic recognition.

[0052] S2. Perform structural semantic recognition on the panoramic image, and divide the panoramic image into multiple regions with structural semantic labels based on a semantic segmentation network. Each semantic label corresponds to a specific structural part of the ship's hull.

[0053] To achieve location-based painting strategy matching, it is essential to identify the corresponding hull structural components within each image region of the panoramic image, establishing a correspondence between image data and structural semantics. Due to the highly complex internal and external structures of ships, including different parts such as roofs, bulkheads, bilges, structural stiffeners, decks, and ventilation shafts, without structural semantic recognition, it is impossible to determine the painting attributes of each image segment, nor can it support subsequent painting type inference and quality standard loading for local images. Therefore, it is necessary to perform structural-level partitioning of the panoramic image based on semantic segmentation technology and assign corresponding structural semantic labels to each region to complete the encoding of structural information at the image level.

[0054] In this step, structural semantic recognition refers to the process of classifying pixels in the input image according to their function or physical location in the ship's structure, and outputting a structural semantic map, in which each pixel is assigned a semantic label.

[0055] Semantic tags are identification information used to represent the structural category to which a certain area in an image belongs. Common tags include top, bottom, longitudinal bulkhead, transverse bulkhead, stiffener, deck, ventilation duct, pipe support, etc.

[0056] like Figure 2As shown, the specific implementation process of step S2 is as follows: The panoramic unfolded image obtained in step S1 is used as input image data and input into a pre-trained semantic segmentation neural network model. Optional implementation schemes for the semantic segmentation neural network include: a spherical convolutional network structure modified from DeepLabv3, a segmentation model using Transformers structures such as SegFormer, or a dual-branch multi-scale network integrating convolution and attention mechanisms. In the specific implementation, the image is first normalized in size and an image tensor is constructed as input. After forward propagation of the network, a structural semantic map tensor is output, with the same size as the original image, and each pixel position contains a classification probability vector. The system selects the classification result according to the maximum probability principle and generates the final structural semantic label map.

[0057] To improve the accuracy and generalization ability of semantic recognition, a dataset with multiple categories, ship types, and perspectives needs to be introduced as training samples during the model training phase. The dataset can be obtained by manually annotating historical ship hull engineering images and combining BIM or CAD ship structure projection images.

[0058] Preferably, an auxiliary loss function based on structural symmetry and topological constraints can be introduced during the training process to improve the model's ability to identify structural boundaries and slender structures (such as stiffeners and bulkhead edges).

[0059] Ship structures widely employ symmetrical arrangements and regular topological connections in their design. For example, longitudinal ribs and transverse stiffeners are typically evenly distributed or arranged axially symmetrically on bulkheads and bottom plates, exhibiting significant symmetry and connectivity characteristics in their structural morphology. Therefore, if the model relies solely on traditional pixel-level label loss functions during training, it is easy to overlook these structural geometric priors, leading to problems such as blurred boundaries, disconnected connections, and interruptions in slender structures.

[0060] To address this, structural symmetry constraints and topology preservation constraints are introduced as auxiliary loss mechanisms during the model training phase. The structural symmetry constraint, based on the geometric symmetry of the hull structure in the longitudinal or transverse direction, compares the differences in the semantic labels of symmetrical positions in the model's predictions, constructing a symmetry loss function L_sym to ensure consistency in the predictions across the structural symmetry axis. For example, if a section's structure should be symmetrically distributed above and below the centerline, a mirror consistency constraint is applied to that region during training to enhance the model's robustness in recognizing symmetrical structures.

[0061] Topological constraints are used to maintain the connectivity, integrity, and spatial consistency of the predicted structural outline. For slender, linear structures such as stiffeners, frames, and bulkhead edges, unconstrained models are prone to prediction errors such as breaks, discontinuities, or morphological degradation. To address this, a structural connectivity loss L_topo can be constructed based on graph topology theory. For example, the Euler number (the difference between the number of connected components and the number of holes) or a connectivity index based on the edge gradient graph can be used as a supervision criterion to guide the model to generate structural regions with consistent morphology and complete connectivity during prediction.

[0062] In the specific implementation, the total loss function L_total is defined as:

[0063] L_total=L_main+λ1×L_sym+λ2×L_topo

[0064] in:

[0065] L_main represents the basic semantic segmentation loss function, such as cross-entropy loss or Dice coefficient loss;

[0066] L_sym represents the structural symmetry loss function, which can be defined as the mean squared error or structural similarity measure of the difference between the pixel labels on the axis of symmetry between the predicted result and its mirror image.

[0067] L_topo represents the topology loss function, which can be constructed based on metrics such as shape preservation of the structural outline, regional connectivity, and shortest path continuity.

[0068] λ1 and λ2 are weighting adjustment coefficients used to control the contribution of each auxiliary term to the total loss.

[0069] To improve efficiency and deployability, the structural symmetry axis can be extracted from the CAD model or determined by known design parameters. Topological constraints can be obtained by post-processing the prediction graph to obtain the structural edge graph and then calculating the connectivity index. The auxiliary loss can be calculated in parallel with the main loss and fused layer by layer in the multi-scale feature graph.

[0070] In terms of topological constraints, topology-aware modules based on Morse complexes or graph convolutional networks can be used to dynamically capture the connected structures in images, further enhancing the ability to preserve complex structural morphologies.

[0071] For example, when performing semantic segmentation on the internal structure of a large cargo ship's ballast tank, using a basic pixel classification model can easily lead to problems such as broken bulkhead stiffeners and blurred edges on support beams. By introducing the aforementioned symmetry and topology losses, the model can automatically identify symmetrical ribs and maintain their continuous structural boundaries, making the prediction results more consistent with the actual engineering structural layout and improving the accuracy of subsequent painting area identification and process matching.

[0072] In a preferred implementation, the model training employs a perceptual loss based on symmetry structure awareness. Its core principle is to guide the model to explicitly learn and maintain the semantic consistency of symmetrical structures during training by measuring the similarity between the predicted image and its mirror image along the structural symmetry axis in a high-level semantic feature space. This method overcomes the limitations of traditional pixel-level consistency constraints by using feature activation extracted from intermediate layers of the neural network to represent the abstract semantics of the hull structure, thereby achieving structural-level symmetry awareness. It is particularly suitable for semantic consistency recognition of repetitive structural regions such as symmetrical ribs, support ribs, and frame structures in images.

[0073] Ship hull design exhibits a high degree of regularity, particularly in the structural arrangement along the longitudinal and transverse central axes, often displaying a clear geometric symmetry. For example, cargo holds, ballast tanks, and bulkhead support components are typically arranged symmetrically on either side of the central axis. Traditional image semantic segmentation models, based on local texture, edge, and color features, struggle to capture this overall structural regularity, resulting in inconsistent labeling of symmetrical regions, edge jumps, or omissions. Perceptual loss, however, by comparing intermediate features from a pre-trained neural network, can capture semantic features such as shape, boundaries, and distribution at the structural level, enabling the model to learn the deep-level relationships between paired components.

[0074] Specifically, firstly, for each input image prediction result I_pred, a mirroring operation is performed based on a predefined structural symmetry axis (e.g., the longitudinal centerline or the transverse axis) to obtain a mirrored image I_mirror. The mirroring method depends on the direction of the structural symmetry axis; if the hull's symmetry axis is vertical, a horizontal flip is performed; if the symmetry axis is horizontal, a vertical flip is performed. This process can be accomplished using an affine transformation matrix, ensuring consistency of coordinates, scale, and attitude in physical space.

[0075] Secondly, an image classification network pre-trained on a general dataset (such as ImageNet) is used as a perceptual feature extractor. Common networks include VGG16, ResNet50, and EfficientNet. During training, I_pred and I_mirror are input into this network, and its intermediate feature layer outputs F_l(I_pred) and F_l(I_mirror) are extracted respectively, where l represents the convolutional layer level, such as conv3, conv4, conv5, etc.

[0076] Perceptual symmetry loss is defined as the Euclidean distance between the selected feature layers:

[0077] L_sym_perceptual=Σ(w_l×||F_l(I_pred)-F_l(I_mirror)||2)

[0078] Where L_sym_perceptual represents the symmetry loss of the structural semantic layer, F_l represents the feature map output by the l-th convolutional network, w_l is the feature layer weight, and ||·||2 represents the L2 norm. The introduction of multi-layer features allows the loss function to simultaneously focus on local structural details and global spatial symmetry.

[0079] The perceptual symmetry loss function is combined with the main segmentation loss function to form the total loss function:

[0080] L_total=L_main+λ×L_sym_perceptual

[0081] Where L_main is the main semantic segmentation loss (such as cross-entropy), and λ is a parameter that adjusts the weights of the perceptual symmetry loss, usually ranging from 0.05 to 0.2.

[0082] To avoid excessive mirror differences affecting model stability in the early stages of training, a gradual enhancement strategy can be adopted, which involves training only the main loss in the first few rounds of the model, introducing a symmetric loss in the early stages of convergence and gradually increasing its weight; or a warm-up strategy can be used to dynamically adjust λ.

[0083] In a preferred implementation, to improve the model's ability to perceive the symmetry constraints of slender components, especially the segmentation effect of slender strip-shaped regions with obvious directionality such as longitudinal stiffeners, transverse frames, and bulkhead edges commonly found in ship structures, the system can introduce a direction-aware convolution kernel during the construction of the perception loss to perform structural orientation-enhanced feature extraction on the mirror image.

[0084] Direction-aware convolutional kernels are structures designed with direction selectivity on top of traditional convolutional kernels, exhibiting stronger responsiveness in specific directions. For example, for longitudinal stiffeners in a ship's hull, a convolutional kernel with a large receptive field in the vertical direction can be constructed to enhance the recognition of the continuity of longitudinal extensions; for the lateral boundary structures of bulkhead edges, a convolutional kernel with enhanced lateral response can be used to improve the model's discrimination ability in that direction. These convolutional kernels can be hard-coded using predefined directional structures or dynamically generated convolutional weights with directional biases through guided training.

[0085] In the specific implementation process, during the feature extraction stage of the perceptual loss, the mirrored image is input into a neural network model containing orientation-aware convolutional layers to extract orientation-enhanced feature maps at specific levels. These maps are then compared with the orientation-enhanced features extracted from the original predicted image to construct the orientation-aware perceptual loss function. This loss term, while preserving the semantic symmetry of the structure, further emphasizes the coherence and consistency of the structure in a specific orientation, effectively constraining the model to avoid label jumps or boundary breaks caused by small scale, weak texture, or local occlusion when processing slender components.

[0086] In addition, to further enhance the orientation perception capability, the system can also integrate orientation gradient features (such as the orientation gradient map extracted by the Sobel filter) as an additional channel input to the perception network, so that the model has orientation perception capability in the low-level feature extraction stage, thereby improving the ability to abstractly express the symmetric relationship between components in different directions.

[0087] By introducing orientation-aware convolutional kernels, the model not only possesses the ability to understand the symmetry of macroscopic structures but also significantly improves the semantic consistency of complex components such as slender, thin-walled, and sequential structures at the detail level. This addresses common issues in traditional semantic segmentation models for detecting narrow-band ship structures, such as blurred boundaries, distorted shapes, and discontinuous labels, thus providing a more accurate structural segmentation foundation for anti-corrosion coating quality inspection. Through the aforementioned perceptual symmetry loss mechanism, the model not only learns the consistency of the image's label space but also learns the overall symmetry rules of the ship structure in the deep semantic space. This method significantly improves the model's robustness to symmetrical structures such as bulkhead symmetrical components, rib sequences, and frame partitions, reducing misclassification and structural interruption caused by local texture differences. In particular, it exhibits superior boundary integrity and label consistency in complex structural regions, providing a reliable structural semantic foundation for subsequent fine segmentation of the coating area and structure-process matching.

[0088] In a preferred implementation, to avoid label drift caused by confusion between adjacent regions in the image, regional connectivity analysis can be performed on the semantic segmentation results. The maximum connected component extraction algorithm and boundary shape constraint rules are then used to post-process and correct the preliminary recognition results, so that the semantic label boundary is consistent with the actual structural shape.

[0089] In the aforementioned continuation example, structural semantic recognition is performed on the panoramic image of the ballast tank interior captured and stitched by Insta360Pro2. The system uses a pre-trained SegFormer-B5 network model as the segmentation backbone network to identify five structural regions on the image: the roof, longitudinal bulkheads, floor, ribs, and stairways. Each region is assigned a corresponding semantic label, providing a structural basis for subsequent location-matching-based painting strategies. The final output is a structural semantic image of the same size as the input image, where different colors represent different structural parts, serving as an index for local image structure inversion in the next step.

[0090] S3. Based on the structural semantic tags, establish structure-spraying process mapping rules to form a mapping relationship between structure type and corresponding spraying type.

[0091] To achieve intelligent judgment of the spraying process based on structure, it is necessary to establish a mapping relationship between the identified structural semantic tags and the corresponding anti-corrosion spraying process specifications. Due to the different environments, stress states, and functions of different structural parts of a ship, the applicable anti-corrosion spraying types also vary significantly. For example, areas below the waterline require antifouling coatings with resistance to seawater immersion and biofouling, while the interior of ballast tanks requires thick-film epoxy paints with strong adhesion and high pressure resistance. Therefore, by establishing mapping rules between structure and spraying, the automatic conversion of structural semantic tags into spraying detection parameters can be achieved, supporting the determination of spraying type in local images and the subsequent loading of quality inspection strategies.

[0092] The coating type refers to the anti-corrosion coating system specified for a specific structural part, including the type of primer, intermediate coating and topcoat, as well as the target thickness, anti-corrosion performance requirements and testing standards of each coating layer.

[0093] The implementation process of step S3 includes the following operations: First, based on existing classification society standards, engineering specifications, and historical spraying process data, a structure-spraying mapping rule base is constructed. The rule base includes information such as structural semantic tags, recommended spraying systems, target film thickness ranges, typical defect types to be detected, and acceptance thresholds. Optional implementation schemes include: establishing a static dictionary mapping between structural tags and spraying strategies using key-value pairs, or using a decision tree-based set of logical rules to implement reasoning from structure to spraying parameters.

[0094] In practical implementation, the system receives a structural semantic map as input, calls the mapping rule library for each labeled region, and outputs the corresponding spraying type. Preferably, the mapping relationship can be designed as a multi-level rule chain, for example, first matching the structural category (such as "cabin interior"), then refining it to specific parts (such as "ballast bulkhead"), and further matching the spraying environment (such as "normal temperature construction" or "humid and highly corrosive area") to select the optimal spraying system.

[0095] In a preferred implementation, to accommodate the differences in painting strategies between different ship types and projects, the structure-painting mapping rule base can be designed as a configurable JSON structure or a database table structure, supporting manual editing and version management, and allowing automatic synchronization through interface integration with the shipyard's design and process management system.

[0096] If we let the set of structural semantics be S and the set of spraying types be T, then the structure-spraying mapping relationship can be expressed as a mapping function:

[0097] T(s)=f(s)

[0098] Where s represents the structural semantic label, T(s) represents the corresponding spraying type, and f represents the decision function that maps the structural label to the spraying rule. Specifically, it can be a lookup table mapping, a conditional inference function, or a machine learning classifier.

[0099] Continuing with the previous example, after performing semantic recognition on the panoramic image of the ballast tanks of a bulk carrier, the region was identified as "longitudinal bulkhead." The system, after querying the rule base, determined that the structure should use a "thick-film epoxy system," with a solvent-free epoxy primer, a target thickness of 300 to 350 micrometers, and that quality inspection should focus on three defect indicators: blistering, peeling, and abnormal adhesion. This coating type information is then used as the coating parameter input in subsequent local image detection.

[0100] S4. Acquire local images: Use handheld devices, mobile terminals, or inspection robots to acquire local images of the area where the ship hull has been painted.

[0101] To achieve detailed quality analysis of the painted areas on the hull, image acquisition of localized areas is necessary after the painting process. Since painting defects such as blistering, missed areas, and runs typically occur at a microscopic scale or in localized areas, they cannot be directly identified from panoramic images. Therefore, high-resolution image acquisition of key areas is required to support subsequent defect detection, film thickness estimation, and surface condition assessment. Localized image acquisition not only provides detailed information but also serves as input for semantic pairing with the panoramic image structure, forming an intermediary bridge between structural localization and painting inspection.

[0102] Localized images refer to high-resolution image data acquired within a localized area, typically with a resolution better than 0.2 millimeters per pixel, clearly reflecting surface texture, color uniformity, and subtle imperfections. Handheld devices refer to image acquisition equipment operated by a human, such as industrial cameras, mobile terminals, or structured light scanners. Inspection robots are robotic platforms capable of moving on ship surfaces or autonomously inspecting cabins, carrying image sensors to perform image acquisition tasks.

[0103] The specific implementation process of step S4 is as follows: After the spraying operation is completed, local images are acquired using an image acquisition device according to the preset detection path or the position indicated by the structural semantic map. Optional implementation schemes include: having an inspection personnel hold a high-definition camera or a terminal device equipped with a halo to take pictures in the target area; or using a tracked, magnetic, or aerial inspection robot equipped with a vision module to complete image acquisition during the structural surface inspection. During the acquisition process, to ensure image quality, the imaging distance should be controlled within a preset range (e.g., 30 to 70 cm), and the vertical viewing angle should be kept stable to avoid image distortion. Preferably, to enhance image contrast and detail visibility, the acquisition system can be equipped with an auxiliary lighting system or use HDR image synthesis technology to adapt to insufficient lighting inside the hull or excessive surface reflection.

[0104] In a preferred implementation, to improve acquisition efficiency and avoid repeated acquisition, the local image acquisition system can integrate a pose perception module or a QR code marking and scanning module to record the reference position and shooting angle of each image in the panoramic view, providing auxiliary data for subsequent spatial back projection and structural annotation.

[0105] If we denote the local image as I_local and the pose during capture as P_local, the acquisition process satisfies the following spatial constraints:

[0106] I_local = Capture(P_local, L)

[0107] Where I_local represents the acquired local image, P_local represents the spatial pose parameters of the shooting device, including the position vector and orientation angle, L represents the illumination intensity and image contrast parameters of the shooting area, and Capture represents the image acquisition function.

[0108] Continuing with the previous example, after the thick-film epoxy paint was applied to the ballast tank, the inspectors used a handheld industrial camera equipped with an LED ring light source to take pictures of different height areas of the bulkhead. Three images were collected for each bulkhead area, with a resolution of 4096×3072.

[0109] S5. Perform spatial location inversion on the local image, match the feature information of the local image with the panoramic image, determine the mapping position of the local image in the panoramic image, and obtain the corresponding structural semantic label.

[0110] To determine the structural attribution of a localized sprayed image, the image must be spatially repositioned to the corresponding hull structural region. Since localized images only cover a limited viewpoint and cannot directly represent the structural location, and the determination of the spraying process highly depends on the semantic information of the structural parts, it is necessary to spatially match the localized image with a pre-constructed panoramic structural semantic map to obtain its projection position in the panoramic image and its corresponding structural semantic label. This operation provides a locational prior for the entire spraying quality inspection process, giving defect detection a structural contextual basis.

[0111] Spatial location delocalization refers to the process of finding the best matching region of a local image within a panoramic image. Its core lies in finding a unique or optimal corresponding position for the visual features of the local image on the panoramic image. The mapped position refers to the projection area of ​​the local image onto the coordinate system of the panoramic image, which can be represented as a two-dimensional coordinate position or a spatial coordinate range.

[0112] like Figure 3As shown, the specific implementation process of step S5 is as follows: First, extract image features from the local image. These features include, but are not limited to, color histograms, gradient orientation histograms, SIFT local feature point sets, ORB keypoint sets, or semantic feature vectors output by intermediate layers of deep networks. Then, match and calculate the similarity between this feature set and the feature sets of each candidate region in the panoramic image to find the best matching region. Optional implementation schemes include: using a brute-force sliding window matching method to traverse and match all locations in the panoramic image, or using a KD-tree-accelerated approximate nearest neighbor algorithm to quickly match feature vectors. Preferably, locality-sensitive hashing or a multi-scale comparison mechanism based on feature pyramids can be introduced to improve matching speed and robustness.

[0113] After matching is completed, the system obtains the structural label of the mapped location in the panoramic structural semantic map, thereby determining which hull structure region the local image belongs to.

[0114] In a preferred implementation, to improve the anti-positioning accuracy of local images in panoramic images, the local image acquisition device may also carry a pose sensing module, such as an inertial measurement unit (IMU), a QR code scanning system, a visual-inertial positioning module (VIO), or other sensing components capable of acquiring displacement and orientation. This pose sensing module can acquire the three-dimensional spatial position information and attitude angles (including pitch, roll, and yaw angles) of the acquisition device in real time at the moment of capture, thereby providing prior parameter constraints for spatial matching between local and panoramic images, narrowing the search range, reducing matching ambiguity, and improving the uniqueness and stability of anti-positioning.

[0115] Specifically, when a local image is acquired, the IMU can simultaneously record acceleration and angular velocity data. After Kalman filtering or sliding window optimization, the device's three-dimensional trajectory and attitude information can be calculated. This information can be initially aligned with the three-dimensional spherical coordinate system or projected coordinate system established by the panoramic image, serving as the initial spatial constraint value for image matching. On the other hand, the QR code scanning system can quickly anchor a local image to a predefined reference coordinate system or calibration point by recognizing the preset QR code icon number and relative position inside the hull. This allows for locating the approximate area within a large panoramic image, and then achieving precise mapping through image content detail matching.

[0116] In practical implementation, the pose data can be input as a parameter to the image feature matching module to filter and weight the spatial consistency of matching point pairs. For example, while calculating the similarity of image feature descriptors, a pose consistency scoring term is introduced, so that candidate regions with similar viewpoints and spatial positions obtain higher matching confidence. In addition, if the system has a time synchronization mechanism, the pose trajectory and the panoramic shooting path can be jointly optimized to further improve the global consistency of local image localization.

[0117] This mechanism is particularly suitable for enclosed compartments inside the hull or areas with highly repetitive structures. Under conditions of lack of texture, large changes in lighting, or high similarity of images in multiple areas, it can significantly reduce the mismatch rate, ensure the uniqueness and stability of the positioning results, and provide a reliable positional basis for subsequent painting process judgment and quality analysis.

[0118] If we let the feature vector of the local image be F_local, and the feature vector of the panoramic image at coordinate position (x,y) be F_pano(x,y), then its matching position (x_hat,y_hat) satisfies the following formula:

[0119] (x_hat,y_hat)=argmin_{x,y}D(F_local,F_pano(x,y))

[0120] Where D represents the feature distance metric function, which can be Euclidean distance, cosine distance or Mahalanobis distance, and (x_hat, y_hat) is the position of the optimal matching position in the panoramic coordinate system.

[0121] Continuing with the previous example, the SIFT feature point set is extracted from the bulkhead images collected by the inspection personnel. The system matches the feature library of the ballast tank panoramic image and locates the best matching position at 135 degrees horizontally and 40 degrees vertically in the panoramic image, with the corresponding structural label being "longitudinal bulkhead".

[0122] S6. Based on the structural semantic tags, determine the spraying type and spraying quality inspection standard corresponding to the local image.

[0123] To achieve accurate detection targeting structural attributes, it is necessary to determine the corresponding coating type and coating quality inspection standards based on the structural semantic tags of local images within the panoramic view. Since the types of anti-corrosion coatings, process parameters, and quality assessment standards used for different structural parts of the hull vary significantly, failure to load corresponding inspection rules based on structural attributes may lead to incorrect defect identification or missed detections. Therefore, a parameter matching mechanism driven by structural semantic tags ensures that the inspection process for each local image is process-relevant, has consistent inspection parameters, and is adaptable to evaluation standards.

[0124] The spraying quality inspection standard refers to the threshold values ​​and judgment rules for inspection parameters for a certain type of spraying. It usually includes the target film thickness range, allowable color difference value, bubble area threshold, allowable sag length, etc., and is configured according to the classification society standard or the technical specifications of the paint supplier.

[0125] The specific implementation process of step S6 is as follows: First, the system receives the structural semantic tags obtained in step S5 and calls the corresponding spraying type according to the pre-established structure-spraying mapping relationship library. Then, the system matches the corresponding quality inspection standard according to the spraying type to form a complete set of inspection parameters. The parameter set may include the lower limit of film thickness T_min, the upper limit of film thickness T_max, the color difference tolerance value Delta_E_max, the maximum bubble size A_bubble_max, and the allowable defect pixel rate P_defect_max, etc. Optional implementation schemes include: using a static dictionary structure to establish a direct mapping from structure to inspection parameters, or constructing a rule-based process decision engine.

[0126] Preferably, the testing standard can be customized according to the shipyard project number, construction batch, or paint brand information. The system supports configuring multiple parameter versions and selects the testing standard version applicable to the current image through an automatic indexing mechanism.

[0127] In a preferred implementation, if a local image has a high-confidence structural label but lacks complete back-projection pose information, the system can associate its spatial distribution area in the panoramic image based on semantic labels and adopt a region-level spraying strategy as an alternative, thereby avoiding misjudgment caused by positioning deviation.

[0128] If we define the structural semantic label as S, the spraying type as T, and the detection standard set as Q, then we have:

[0129] T = f(S)

[0130] Q = g(T)

[0131] Where f represents the mapping function from structural semantic labels to spraying types, g represents the query function from spraying types to detection parameter sets, S is the structural label input, T is the spraying type output, and Q is a parameter set containing multiple detection indicators.

[0132] Continuing with the previous example, in step S5, the system has determined the structural semantic label of the local image to be "longitudinal bulkhead." Looking up the coating type in the mapping rule base, it finds it to be a thick-film epoxy system. The system then loads its quality inspection standards: film thickness ranging from 280 to 350 micrometers, allowable color difference value not exceeding 2.0, bubble diameter not exceeding 1 millimeter, and cumulative area not exceeding 0.5% of the image area. This set of parameters serves as the input threshold for the subsequent automatic detection model of the local image, ensuring that the evaluation process conforms to actual engineering standards.

[0133] S7. Perform spraying quality inspection on the local image, use an image recognition-based defect detection model to analyze the quality of the sprayed area in the local image, and output the quality assessment result.

[0134] A spray coating defect detection model is an image recognition algorithm system that takes an image as input and outputs the defect area, type, and degree of determination. It is usually built based on a deep learning model and is equipped with predefined defect categories and detection parameters.

[0135] The specific implementation process of step S7 is as follows: First, the local image is input into the spraying defect detection model, multi-scale features of the image are extracted, and typical spraying defect areas in the image are identified. The defect types include, but are not limited to, missed spraying, sagging, blistering, peeling, insufficient film thickness, and abnormal color difference.

[0136] Optional implementation schemes include: using a pixel-level segmentation model based on the U-Net structure to identify the boundary of the defect region, or using the YOLO series detection model to achieve target-level defect localization, or combining an image regression network to perform pixel-level estimation of the coating thickness.

[0137] The system assesses the compliance of identified defective areas based on the quality inspection standards corresponding to the structure-coating type. For example, if the detected bubble area exceeds a set threshold or the estimated film thickness is below the lower limit, the image is deemed unqualified, and the corresponding location and defect type are recorded. After inspection, the system outputs quality assessment results, including whether the image is qualified, the type of defect, the coordinates of the defect location, the defect area ratio, the film thickness range, and the color difference value.

[0138] In a preferred implementation, the defect detection model can incorporate a multimodal fusion mechanism, inputting RGB images, infrared thermal images, and structured light depth maps into the model to improve the accuracy of identifying latent defects such as blistering and film thickness fluctuations. Furthermore, to enhance detection efficiency and processing speed, lightweight models such as the MobileNet-DeepLab architecture can be used and deployed on edge computing devices to support real-time on-site detection.

[0139] If we define the image input as I_local and the model output as D_local, then the quality evaluation result Q_eval can be defined by the following formula:

[0140] Q_eval=h(D_local,Q_standard)

[0141] Where I_local represents the local image, D_local is the set of defect information output by the detection model, Q_standard is the set of quality judgment parameters corresponding to the structure-coating standard, and h represents the evaluation function, which compares the detection results with the standard and outputs whether it is qualified and various quantitative indicators.

[0142] The implementation of this step has the following beneficial effects: it realizes the automation and intelligence of the spraying quality inspection process, significantly improves inspection efficiency and coverage, reduces human error and the risk of missed inspection, and has advantages such as structural semantic association, accurate defect identification, and quantitative output of indicators, providing a reliable foundation for closed-loop control of spraying process quality.

[0143] Continuing with the previous example, the system has determined that the local image belongs to the "longitudinal bulkhead" structure, corresponding to a thick-film epoxy coating. The system inputs the image into the U-Net semantic segmentation model deployed on the edge terminal. The detection results show a bubbling area with an area of ​​1.2 square centimeters, an estimated film thickness of 180 micrometers, below the standard lower limit of 280 micrometers, and a color difference value of 1.5, which is within the allowable range. Based on this, the system outputs an inspection report determining that the image is unqualified and marks the defect location for rework or subsequent tracking and verification.

[0144] In another embodiment, the present invention also provides a machine vision-based ship anti-corrosion coating quality inspection system, comprising:

[0145] The panoramic image acquisition module is used to acquire 360-degree panoramic images of the space to be inspected on the ship's hull. It takes panoramic pictures of the ship's external and / or internal structure through panoramic imaging equipment to form panoramic image data unfolded on a spherical surface.

[0146] The structural semantic recognition module is used to perform structural semantic recognition on the panoramic image. Based on the semantic segmentation network, the panoramic image is divided into multiple regions with structural semantic labels, and each semantic label corresponds to a specific structural part of the ship's hull.

[0147] The spraying rule matching module is used to establish structure-spraying process mapping rules based on the structure semantic tags, forming a mapping relationship between structure type and corresponding spraying type;

[0148] The local image acquisition module is used to acquire local images of the painted area of ​​the ship hull using a handheld device, mobile terminal, or inspection robot.

[0149] The spatial location anti-localization module is used to match the feature information of the local image with the panoramic image, determine the mapping position of the local image in the panoramic image, and obtain the corresponding structural semantic label.

[0150] The spraying standard extraction module is used to determine the spraying type and spraying quality inspection standard corresponding to the local image based on the structural semantic tags.

[0151] The quality inspection module is used to inspect the spraying quality of the local image. It uses an image recognition-based defect detection model to analyze the quality of the sprayed area in the local image and outputs the quality assessment results.

[0152] Preferably, the structural semantic recognition module is configured as a semantic segmentation network that incorporates an auxiliary loss function during training. This auxiliary loss function is constructed based on structural symmetry and topological constraints, and the total loss function is:

[0153] L_total=L_main+λ1×L_sym+λ2×L_topo,

[0154] Where L_main represents the basic semantic segmentation loss function, L_sym represents the structural symmetry loss function, L_topo represents the topological structure loss function, and λ1 and λ2 are weight adjustment coefficients.

[0155] Preferably, the structural semantic recognition module further includes a symmetry structure perception module, used to introduce a symmetry perception loss to construct structural semantic symmetry constraints. The symmetry perception loss is constructed as follows:

[0156] For each prediction result I_pred, a corresponding mirror image I_mirror is generated. The features F_l(I_pred) and F_l(I_mirror) of the l-th layer of the pre-trained classification network are extracted, and a perceptual symmetry loss function is constructed.

[0157] L_sym_perceptual=Σ(w_l×||F_l(I_pred)-F_l(I_mirror)||2),

[0158] Where w_l is the weighting coefficient of the feature layer, F_l is the feature map of the l-th layer, and ||·||2 is the L2 norm;

[0159] The structural semantic recognition module is further configured to weightedly fuse the loss function with the basic semantic segmentation loss L_main to form the total loss function:

[0160] L_total=L_main+λ×L_sym_perceptual,

[0161] Where λ is the weighting adjustment coefficient.

[0162] Preferably, the local image acquisition module includes a pose sensing module that works synchronously with it. The pose sensing module is one or a combination of an IMU (Inertial Measurement Unit), a QR code recognition system, or a visual-inertial positioning module.

[0163] The pose perception module is used to simultaneously record acceleration and angular velocity data when acquiring local images, and optimize the three-dimensional trajectory and attitude information of the computing device through Kalman filtering or sliding window as the initial value for spatial anti-positioning.

[0164] Preferably, the spatial location anti-positioning module is configured to achieve location matching in the following manner:

[0165] The feature vector F_local of the local image is matched with the feature vector F_pano(x,y) of the panoramic image at coordinate position (x,y) to determine the matching position (x_hat, y_hat), satisfying the following relationship:

[0166] (x_hat,y_hat)=argmin_{x,y}D(F_local,F_pano(x,y)),

[0167] Wherein, D is the feature distance metric function, which is one of Euclidean distance, cosine distance or Mahalanobis distance.

[0168] It should be noted that the explanation of the aforementioned embodiment of the machine vision-based ship anti-corrosion spraying quality inspection method also applies to the apparatus of the embodiments of this application, and will not be repeated here.

[0169] Those skilled in the art will recognize that the units and algorithm steps described in the embodiments disclosed herein can be implemented using electronic hardware, computer software, or a combination of electronic hardware and software. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0170] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0171] In the several embodiments provided in this application, any function, if implemented as a software functional unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0172] The above description is merely a specific embodiment of this application. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the protection scope of this application. For some module structures not specifically defined in this invention, the content described in the prior art shall prevail. The prior art mentioned in the foregoing background and specific embodiments sections can be considered as part of this invention and used to understand the meaning of some technical features or parameters.

Claims

1. A machine vision-based method for inspecting the quality of ship anti-corrosion spraying, characterized in that, The method includes the following steps: S1. Obtain a 360-degree panoramic image of the space to be inspected on the hull. Take panoramic photos of the external and / or internal structure of the ship using a panoramic imaging device to form panoramic image data unfolded on a spherical surface. S2. Perform structural semantic recognition on the panoramic image, and divide the panoramic image into multiple regions with structural semantic labels based on a semantic segmentation network. Each semantic label corresponds to a specific structural part of the ship's hull. S3. Based on the structural semantic tags, establish structure-spraying process mapping rules to form a mapping relationship between structure type and corresponding spraying type; S4. Acquire local images by using handheld devices, mobile terminals, or inspection robots to collect local images of the area where the ship hull has been painted. S5. Perform spatial location inversion on the local image, match the feature information of the local image with the panoramic image, determine the mapping position of the local image in the panoramic image, and obtain the corresponding structural semantic label. S6. Based on the structural semantic tags, determine the spraying type and spraying quality inspection standard corresponding to the local image; S7. Perform spraying quality inspection on the local image, use an image recognition-based defect detection model to analyze the quality of the sprayed area in the local image, and output the quality assessment result.

2. The method for inspecting the quality of ship anti-corrosion spraying based on machine vision according to claim 1, characterized in that, The semantic segmentation network uses an auxiliary loss function based on structural symmetry and topological constraints during training.

3. The method for inspecting the quality of ship anti-corrosion spraying based on machine vision according to claim 1, characterized in that, During training, the semantic segmentation network uses a perceptual loss based on symmetry structure awareness. Specifically, the perceptual loss based on symmetry structure awareness is as follows: For each input image prediction result I_pred, a mirror image I_mirror is obtained based on the predefined structural symmetry axis. An image classification network pre-trained on a general dataset is used as a perceptual feature extractor to extract the output of its intermediate feature layers. Perceptual symmetry loss is defined as the Euclidean distance between selected feature layers; The perceptual symmetry loss function is combined with the main segmentation loss function to form the total loss function.

4. The method for inspecting the quality of ship anti-corrosion spraying based on machine vision according to claim 1, characterized in that: When a local image is acquired, the IMU simultaneously records acceleration and angular velocity data. After Kalman filtering or sliding window optimization, the device's three-dimensional trajectory and attitude information are calculated. Subsequently, it is initially aligned with the three-dimensional spherical coordinate system or projected coordinate system established by the panoramic image, serving as the initial value of the spatial constraints for image matching.

5. The method for inspecting the quality of ship anti-corrosion spraying based on machine vision according to claim 1, characterized in that... The step of performing spatial location inversion on the local image includes: The feature vector of the local image is F_local, and the feature vector of the panoramic image at coordinate position (x,y) is F_pano(x,y).

6. A machine vision-based system for inspecting the quality of anti-corrosion spraying on ships, characterized in that, The system includes the following modules: The panoramic image acquisition module is used to acquire 360-degree panoramic images of the space to be inspected on the ship's hull. It takes panoramic pictures of the ship's external and / or internal structure through panoramic imaging equipment to form panoramic image data unfolded on a spherical surface. The structural semantic recognition module is used to perform structural semantic recognition on the panoramic image. Based on the semantic segmentation network, the panoramic image is divided into multiple regions with structural semantic labels, and each semantic label corresponds to a specific structural part of the ship's hull. The spraying rule matching module is used to establish structure-spraying process mapping rules based on the structure semantic tags, forming a mapping relationship between structure type and corresponding spraying type; The local image acquisition module is used to acquire local images of the painted area of ​​the ship hull using a handheld device, mobile terminal, or inspection robot. The spatial location anti-localization module is used to match the feature information of the local image with the panoramic image, determine the mapping position of the local image in the panoramic image, and obtain the corresponding structural semantic label. The spraying standard extraction module is used to determine the spraying type and spraying quality inspection standard corresponding to the local image based on the structural semantic tags. The quality inspection module is used to inspect the spraying quality of the local image. It uses an image recognition-based defect detection model to analyze the quality of the sprayed area in the local image and outputs the quality assessment results.

7. The machine vision-based ship anti-corrosion spraying quality inspection system according to claim 6, characterized in that, The structural semantic recognition module is configured to use a semantic segmentation network that incorporates an auxiliary loss function during training. The auxiliary loss function is constructed based on structural symmetry and topological constraints.

8. The machine vision-based ship anti-corrosion spraying quality inspection system according to claim 6, characterized in that, The structural semantic recognition module further includes a symmetry structure perception module, used to introduce a symmetry perception loss to construct structural semantic symmetry constraints. The symmetry perception loss is constructed as follows: For each prediction result I_pred, a corresponding mirror image I_mirror is generated. The features F_l(I_pred) and F_l(I_mirror) of the l-th layer of the pre-trained classification network are extracted, and a perceptual symmetry loss function is constructed. The structural semantic recognition module is further configured to weightedly fuse the loss function with the basic semantic segmentation loss L_main to form the total loss function.

9. The machine vision-based ship anti-corrosion spraying quality inspection system according to claim 6, characterized in that, The local image acquisition module includes a pose sensing module that works synchronously with it. The pose sensing module is one or a combination of an IMU (Inertial Measurement Unit), a QR code recognition system, or a visual-inertial positioning module. The pose perception module is used to simultaneously record acceleration and angular velocity data when acquiring local images, and optimize the three-dimensional trajectory and attitude information of the computing device through Kalman filtering or sliding window as the initial value for spatial anti-positioning.

10. The machine vision-based ship anti-corrosion spraying quality inspection system according to claim 6, characterized in that, The spatial location anti-positioning module is configured to achieve location matching in the following manner: The feature vector F_local of the local image is matched with the feature vector F_pano(x,y) of the panoramic image coordinate position (x,y) to determine the matching position (x_hat,y_hat).

Images