Ship anticorrosion spraying quality detection method and system based on machine vision

Abstract

The application belongs to the field of image processing, and provides a ship anticorrosion spraying quality detection method and system based on machine vision, which comprises the following steps: acquiring a 360-degree panoramic image of a ship body space to be detected; performing structural semantic recognition on the panoramic image, dividing the panoramic image into multiple regions with structural semantic labels based on a semantic segmentation network; establishing a structure-spraying process mapping rule to form a mapping relationship between structure types and corresponding spraying types; collecting local images of the ship body spraying completed regions; performing spatial position reverse positioning on the local images and acquiring corresponding structural semantic labels; determining the spraying type and spraying quality detection standard corresponding to the local images according to the structural semantic labels; and performing spraying quality detection on the local images, adopting a defect detection model based on image recognition to analyze the quality of the spraying regions in the local images and outputting a quality evaluation result.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of image processing, and specifically relates to a ship anticorrosion spraying quality detection method and system based on machine vision. BACKGROUND

[0002] Ship anticorrosion spraying refers to applying a coating system with anticorrosion function on the surface of the ship steel structure to resist the corrosion damage caused by seawater, heat and humidity, ultraviolet rays, and marine biological attachment, so as to prolong the service life of the ship and ensure the safety of the structure. The coating system is usually composed of primer, intermediate coating, and topcoat, and different parts are configured with different types of coating materials and thickness according to the function and environmental difference, such as zinc-rich primer, thick film epoxy paint, or antifouling paint.

[0003] Due to the complex and diverse ship structure, the spraying area is widely distributed and the environment is significantly different, and the anticorrosion spraying requirements of different parts are also different. For example, the planking above the waterline is exposed to wind, rain, and sunlight, and requires a topcoat with strong ultraviolet resistance and color durability; the area below the waterline requires long-term resistance to seawater immersion and microbial attachment, and must use a coating system with antifouling ability; the internal structure of the ballast tank and cargo hold has strict requirements on adhesion, wear resistance, and impermeability. Therefore, the spraying strategy needs to be customized according to the specific structure position.

[0004] The traditional manual sampling method cannot realize large-area and high-frequency spraying quality evaluation, and relies on experience judgment, which is subjective and low in efficiency, and is easy to miss or misjudge. The ship anticorrosion spraying quality detection method based on machine vision can realize automatic, digital, and objective detection of the spraying area, and through image recognition and intelligent analysis, it can judge whether the coating has defects such as missing spraying, blistering, sagging, and color difference, greatly improving the detection coverage and accuracy, and is an important supporting technology for realizing intelligent shipbuilding and automatic control of shipyards.

[0005] However, this method still faces significant difficulties in engineering implementation. On the one hand, using a large range of images (such as panoramic images or large-format photos) can cover a wide area, but due to the resolution limitation, it is difficult to clearly identify the details of the coating defects; on the other hand, using local images can help with detailed quality detection, but the ship structure is complex, it is difficult to extract semantic structures, and isolated images are difficult to automatically determine their belonging structure area and spraying type, which requires manual annotation or auxiliary information, which is time-consuming and low in efficiency. SUMMARY

[0006] In order to solve the problems in the prior art, the present application provides a ship anticorrosion spraying quality detection method based on machine vision, comprising the following steps:

[0007] S1, acquire a 360-degree panoramic image of a hull space to be detected, panoramic imaging equipment is used to take panoramic pictures of the external and / or internal structure of the ship, and panoramic image data is formed in spherical development;

[0008] S2, structure semantic recognition is performed on the panoramic image, the panoramic image is divided into multiple regions with structure semantic labels based on a semantic segmentation network, and each semantic label corresponds to a specific structural part of the hull;

[0009] S3, a structure-spraying process mapping rule is established according to the structure semantic labels, and a mapping relationship between the structure type and the corresponding spraying type is formed;

[0010] S4, a local image is acquired, and a handheld device, a mobile terminal or a detection robot is used to collect a local image of a hull spraying completed area;

[0011] S5, spatial position reverse positioning is performed on the local image, feature information of the local image is matched with the panoramic image, a mapping position of the local image in the panoramic image is determined, and a corresponding structure semantic label is acquired;

[0012] S6, the spraying type corresponding to the local image and the spraying quality detection standard are determined according to the structure semantic label;

[0013] S7, spraying quality detection is performed on the local image, a defect detection model based on image recognition is used to analyze the quality of the spraying area in the local image, and a quality evaluation result is output.

[0014] Another aspect of the present application also provides a ship anticorrosion spraying quality detection system based on machine vision, comprising the following modules:

[0015] A panoramic image acquisition module is used to acquire a 360-degree panoramic image of a hull space to be detected, panoramic imaging equipment is used to take panoramic pictures of the external and / or internal structure of the ship, and panoramic image data is formed in spherical development;

[0016] A structure semantic recognition module is used to perform structure semantic recognition on the panoramic image, the panoramic image is divided into multiple regions with structure semantic labels based on a semantic segmentation network, and each semantic label corresponds to a specific structural part of the hull;

[0017] A spraying rule matching module is used to establish a structure-spraying process mapping rule according to the structure semantic labels, and a mapping relationship between the structure type and the corresponding spraying type is formed;

[0018] A local image acquisition module is used to collect a local image of a hull spraying completed area by using a handheld device, a mobile terminal or a detection robot;

[0019] The spatial position reverse positioning module is configured to match feature information of the local image with the panoramic image, determine a mapping position of the local image in the panoramic image, and obtain a corresponding structure semantic label;

[0020] The spraying standard extraction module is configured to determine a spraying type and a spraying quality detection standard corresponding to the local image according to the structure semantic label.

[0021] The quality detection module is configured to perform spraying quality detection on the local image, perform quality analysis on a spraying area in the local image by using a defect detection model based on image recognition, and output a quality evaluation result.

[0022] Further, the local image acquisition module comprises a pose perception module that works synchronously therewith, the pose perception module being one of an IMU (inertial measurement unit), a two-dimensional code recognition system or a visual-inertial combined positioning module, or a combination thereof.

[0023] The pose perception module is configured to record acceleration and angular velocity data synchronously when the local image is acquired, and calculate three-dimensional trajectory and attitude information of the device by Kalman filtering or sliding window optimization, as an initial value for spatial reverse positioning.

[0024] The present application fuses information of panoramic images and local images, constructs a mapping model of structure semantics and spraying strategies, and realizes spraying detection process control based on structure position driving. The structure recognition accuracy is improved by symmetric loss, and the structure part and spraying type to which the local image belongs can be automatically inferred by the panoramic structure recognition and spatial reverse positioning mechanism, so that the detection parameters are loaded accordingly, and the engineering adaptability of the detection result is improved.

[0025] The present application adopts a defect recognition model based on deep learning to determine spraying defects, dynamically adapts quality evaluation standards in combination with structure semantic labels and standard libraries, can accurately judge the defect position, type and severity of the spraying area, improves the degree of detection automation, and reduces the dependence on manual work.

[0026] The technical solution of the present application supports on-site rapid deployment and edge device operation, has good real-time performance, scalability and universality, is suitable for detection requirements of different ship types and different spraying systems, and has significant industrial promotion value. BRIEF DESCRIPTION OF DRAWINGS

[0027] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

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

[0029] Figure 2 is the semantic segmentation flowchart of the present application;

[0030] Figure 3 is the anti-positioning flowchart of the present application. DETAILED DESCRIPTION

[0031] The preferred description of the present application is made in combination with the drawings and the detailed description.

[0032] In order to solve the technical problems in the background art, as shown in the background art, the present application provides a ship anticorrosion spraying quality detection method based on machine vision, which belongs to the field of image processing, in order to improve the accuracy and efficiency of quality detection. Figure 1

[0033] The method comprises the following steps:

[0034] S1, acquiring a 360-degree panoramic image of a ship body to be detected, panoramic imaging equipment is used to take panoramic pictures of the external and / or internal structure of the ship, and spherical panoramic image data is formed;

[0035] S2, performing structural semantic recognition on the panoramic image, dividing the panoramic image into multiple regions with structural semantic labels based on a semantic segmentation network, and each semantic label corresponds to a specific structural part of the ship body;

[0036] S3, establishing a structure-spraying process mapping rule according to the structural semantic labels, and forming a mapping relationship between the structure type and the corresponding spraying type;

[0037] S4, acquiring a local image, using a handheld device, a mobile terminal or a detection robot to collect a local image of a ship body spraying completed area;

[0038] S5, performing spatial position anti-positioning on the local image, matching the feature information of the local image with the panoramic image, determining the mapping position of the local image in the panoramic image, and acquiring the corresponding structural semantic label;

[0039] S6, determining the spraying type and spraying quality detection standard corresponding to the local image according to the structural semantic label;

[0040] S7, performing spraying quality detection on the local image, using an image recognition-based defect detection model to analyze the quality of the spraying area in the local image, and outputting a quality evaluation result.

[0041] The above steps will be further explained one by one.

[0042] ​S1, acquire a 360-degree panoramic image of a hull space to be detected, and panoramic imaging equipment is used to take panoramic pictures of the external and / or internal structure of the ship, forming a spherical panoramic image data.

[0043] In order to realize the structural correlation automatic detection of the ship anticorrosion spraying quality, the complete structure information of the space where the hull is located must be collected before the whole detection process starts, and the semantic mapping relationship between the image and the structure is established. Due to the complex structure of the hull, there are large curved surfaces, deep internal cabins, corner dead angles and other areas. A single view image cannot fully cover the target detection space, and the structure position attribute of the subsequent local image cannot be accurately judged. Therefore, the panoramic image of the space to be detected is first acquired, and the structure semantic segmentation of the image is performed, which provides structure prior support for the structure positioning and spraying type determination of the subsequent local image.

[0044] In this step, the panoramic image refers to the image data expressing the full-view information of a space in the form of spherical or cylindrical expansion, which can cover the image information of 360 degrees horizontally and 180 degrees vertically from the current shooting position. The panoramic imaging equipment refers to the equipment that can complete multi-directional synchronous imaging at a single time point or in a short time, and output panoramic images through splicing or reflection expansion. Common devices include spherical cameras, structured light camera arrays, fisheye camera groups, etc.

[0045] The specific implementation process of this step S1 is as follows: first, panoramic imaging equipment is arranged in the detection area to collect the whole scene of the external structure or internal cabin of the hull. In the specific implementation, the optional implementation scheme includes: using a spherical multi-lens camera (such as an omnidirectional camera device composed of six groups of wide-angle cameras) to synchronously collect images in each direction at the same position; or using a rotatable high-resolution single lens with a mechanical turntable for sequential multi-angle shooting, and then generating a panoramic image through image stitching algorithm. The stitching algorithm can be selected from SIFT feature point matching algorithm, RANSAC model registration algorithm, and image re-projection method based on spherical projection. Preferably, to improve the structural consistency and light balance of image acquisition, the equipment can be equipped with an automatic exposure control module and an HDR imaging module to adapt to the weak light environment inside the hull and the strong light reflection environment outside the hull.

[0046] After the image acquisition is completed, the original image is geometrically expanded using a spherical projection model to generate spherical expansion image data. The principle of spherical expansion is to map the spherical space coordinates (θ, φ) to the two-dimensional image coordinates (x, y), and the mapping relationship is:

[0047] x=Wx(θ+π)÷2π

[0048] y=Hx(π÷2-φ)÷π

[0049] Wherein, θ represents the horizontal angle (unit: radian), φ represents the vertical angle (unit: radian), W is the horizontal resolution of the image, H is the vertical resolution of the image, x and y are the corresponding pixel position coordinates in the projection image.

[0050] In a preferred implementation, in order to enhance the stability of subsequent structural semantic recognition, the panoramic image data will be spatially registered according to the preset hull design reference points (such as hatch number, structure number or pre-buried positioning label), so as to realize the spatial back-projection and structural position determination of the local image subsequently.

[0051] In an example, for the cabin anticorrosion detection work of a 200,000-ton bulk carrier, a spherical panoramic camera Insta360Pro2 is installed at the center of each ballast tank, and synchronous shooting at four angles is performed. The images are fused into a panoramic unfolded image with a resolution of 7680x3840 by a deep learning image stitching engine, and then form the input basic image data for structural semantic recognition.

[0052] S2, performing structural semantic recognition on the panoramic image, dividing the panoramic image into multiple regions with structural semantic labels based on a semantic segmentation network, each semantic label corresponding to a specific structural part of the hull.

[0053] In order to realize the matching of spraying strategy based on position association, it is necessary to identify the hull structure part corresponding to each image region in the panoramic image to establish the corresponding relationship between image data and structural semantics. Due to the high complexity of the internal and external structure of the ship, there are different parts such as cabin roof, cabin wall, cabin bottom, structural reinforcement rib, deck, ventilation shaft, etc. If structural semantic recognition is not performed, the spraying properties of each image segment cannot be determined, and the subsequent spraying type reasoning and quality standard loading of local image cannot be supported. Therefore, it is necessary to perform structural level zoning processing on the panoramic image based on semantic segmentation technology, and assign each region with a corresponding structural semantic label to complete the structural information coding 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 position in the hull structure, and the output is a structural semantic map, i.e. each pixel is assigned a semantic label.

[0055] The semantic label refers to the identification information used to represent the structural category to which a region in the image belongs. Common labels include cabin roof, cabin bottom, longitudinal bulkhead, transverse bulkhead, reinforcement rib, deck, ventilation pipe, pipe support, etc.

[0056] For example, the semantic segmentation network used in this example is a DeepLabV3+ network, which is a semantic segmentation network based on the DeepLabV3 network. Figure 2As shown, the specific implementation process of this step S2 is as follows: the panoramic unfolded image obtained in step S1 is taken as input image data and input into the semantic segmentation neural network model which has been pre-trained. The optional implementation scheme of the semantic segmentation neural network includes: a spherical convolution network structure modified by DeepLabv3, a segmentation model with a Transformers structure such as SegFormer, or a double-branch multi-scale network fusing convolution and attention mechanism. In the specific implementation, first, the image is subjected to size normalization processing and an image tensor input is constructed, and after the network forward propagation, a structure semantic graph tensor is output, which has the same size as the original image and contains a classification probability vector at each pixel position. The system selects the classification result according to the maximum probability principle and generates the final structure semantic label graph.

[0057] In order to improve the accuracy and generalization ability of semantic recognition, a multi-class, multi-ship type and multi-view dataset needs to be introduced as a training sample in the model training stage. The dataset can be obtained by manually labeling historical ship body engineering images and combining BIM or CAD ship structure projection images.

[0058] Preferably, an auxiliary loss function based on structural symmetry and topological constraint can be introduced in the training process to improve the recognition ability of the model for structure boundaries and slender structures (such as stiffeners and bulkhead edges).

[0059] Ship structures widely adopt symmetrical arrangement and regular topological connection in design, such as longitudinal ribs and transverse stiffeners which are usually uniformly distributed or arranged in axial symmetry on bulkheads and bottom plates, and the structure morphology has significant symmetry and connectivity characteristics. Therefore, if the model only relies on the traditional pixel-level label loss function during the training process, it is easy to ignore these structural geometric priors, resulting in problems such as blurred boundaries, disconnected connections and interrupted slender structures.

[0060] Therefore, structural symmetry constraints and topological preservation constraints are introduced as auxiliary loss mechanisms in the model training stage. Among them, the structural symmetry constraint is based on the geometric symmetry relationship of the ship body structure in the longitudinal or transverse direction, compares the difference of the semantic label outputs at the symmetric positions in the model prediction result, constructs a symmetry loss function L_sym, and makes the prediction result consistent on both sides of the structural symmetry axis. For example, if the structures on both sides of the centerline of a cabin section should be symmetrically distributed, mirror consistency constraints are applied to this region during the training process to enhance the recognition robustness of the model for symmetric structures.

[0061] Topology constraints are used to maintain the connectivity, integrity and spatial consistency of the predicted structure contours. For elongated and linear structures such as stiffeners, frames, bulkhead edges, etc., the model is prone to generate prediction errors such as breakage, breakpoints or shape degradation without constraints. Therefore, based on graph topology theory, a structure connectivity loss L_topo can be constructed, for example, using 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 map as a supervision standard to guide the model to generate a morphologically consistent and connected structure region during prediction.

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

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

[0064] Wherein:

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

[0066] L_sym represents the structure symmetry loss function, which can be defined as the mean square error of the pixel label difference between the prediction result and its mirror image on the symmetry axis or the structure similarity measure;

[0067] L_topo represents the topological structure loss function, which can be constructed based on morphological preservation indicators, region connectivity, shortest path continuity, etc.

[0068] λ1 and λ2 are weight adjustment coefficients for controlling the contribution of each auxiliary term in the total loss.

[0069] To improve efficiency and deployability, the structure symmetry axis can be extracted from the CAD model or determined by known design parameters, the topology constraint can be calculated by the structure edge map obtained by post-processing the predicted graph, and the auxiliary loss can be calculated in parallel with the main loss and fused layer by layer in the multi-scale feature map.

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

[0071] For example, when performing semantic segmentation on the internal structure of a large cargo ship ballast tank, using a basic pixel classification model, problems such as breakage of bulkhead stiffeners and blurring of support beam edges are prone to occur. After introducing the above symmetry and topology loss, the model can automatically identify symmetric ribs and maintain their continuous structure boundaries, making the prediction results more consistent with the actual engineering structure layout and improving the accuracy of subsequent spray area identification and process matching.

[0072] In a preferred implementation, a perceptual loss with symmetry structure perception is used in model training. The core principle is to guide the model to learn and maintain the semantic consistency of the symmetric structure during the training process by measuring the similarity between the predicted image and its mirror image on the symmetric axis of the structure in the high-level semantic feature space. This method breaks through the limitations of traditional pixel-level consistency constraints and realizes structure-level symmetry perception through abstract semantic representation of the ship structure extracted by intermediate layer features of the neural network, which is particularly suitable for semantic consistency identification of repeated structure areas such as symmetric ribs, support bars, and frame structures in images.

[0073] Ship design has a high degree of regularity, especially in the structure arrangement on both sides of the longitudinal and transverse center axes, which often presents obvious geometric symmetry. For example, cargo holds, ballast tanks, and bulkhead support members are often symmetrically arranged on both sides of the center axis. Traditional image semantic segmentation models are based on local texture, edge, and color features for classification, which are difficult to capture this overall structural regularity, resulting in inconsistent labels, edge jumps, or omissions in symmetric areas. The perceptual loss can capture structure-level semantic features such as shape, boundary, and distribution by comparing pre-trained intermediate features, allowing the model to learn the deep relationship between "paired" components.

[0074] Specifically, first, for each input image prediction result I_pred, a mirror image I_mirror is obtained by performing a mirror operation according to a predefined structural symmetry axis (such as the longitudinal centerline or transverse axis). The mirror method depends on the direction of the structural symmetry axis. If the ship symmetry axis is vertical, perform horizontal flipping; if the symmetry axis is horizontal, perform vertical flipping. This process can be completed through an affine transformation matrix, ensuring the consistency of coordinates, proportions, and poses in the physical space.

[0075] Second, 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 the network to extract intermediate feature layer outputs F_l(I_pred) and F_l(I_mirror), respectively, where l represents the convolution level, such as conv3, conv4, and conv5.

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

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

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

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

[0080] L_total = L_main + λ x L_sym_perceptual

[0081] where L_main is the main semantic segmentation loss (such as cross-entropy), and λ is a parameter for adjusting the weight of the perceptual symmetry loss, usually taking a value between 0.05 and 0.2.

[0082] To avoid the excessive mirror difference at the initial stage of training affecting the stability of the model, a gradual enhancement strategy can be used, that is, only the main loss is trained in the first few rounds of the model, and the symmetry loss is introduced at the initial stage of convergence and gradually increases its weight; or a warm-up strategy is used to dynamically adjust λ.

[0083] In a preferred implementation, to improve the model's ability to perceive the symmetry constraint of slender members, especially the segmentation effect of slender regions with obvious directionality such as longitudinal stiffeners, transverse skeletons, and bulkhead edges commonly found in ship structures, the system can introduce a direction perception convolution kernel in the construction process of the perceptual loss to perform structure direction-enhanced feature extraction on the mirror image.

[0084] The direction perception convolution kernel is a structure designed based on the traditional convolution kernel with direction selectivity, which has stronger response capability in a specific direction. For example, for longitudinal stiffeners in the ship body, a convolution kernel with a large receptive field in the vertical direction can be constructed to enhance the continuity recognition of the longitudinal extension form; for the transverse boundary structure of the bulkhead edge, a convolution kernel with strong transverse response can be used to improve the model's discrimination ability in that direction. The above convolution kernel can be hard-coded by predefining the direction structure, or dynamically generated with direction-biased convolution weights through guided training.

[0085] In the specific implementation process, in the feature extraction stage of the perceptual loss, the mirror image is input into the neural network model containing the direction perception convolution layer, the direction-enhanced feature map of the specific level is extracted, and compared with the direction-enhanced feature extracted from the original prediction image to construct a direction perception loss function. This loss term further emphasizes the continuity and consistency of the specific direction structure while preserving the structural semantic symmetry, effectively constraining the model to avoid label jumping or boundary fracture when processing slender members due to small scale, weak texture, or local occlusion.

[0086] In addition, to further enhance the direction perception ability, the system can also fuse the direction gradient feature (such as the direction gradient map extracted by the Sobel filter) as an additional channel input into the perception network, so that the model has direction perception ability in the low-level feature extraction stage, thereby improving the abstract expression ability of the symmetry relationship between different direction components.

[0087] By introducing the direction perception convolution kernel, the model not only has the understanding ability of the symmetry of the macro structure, but also significantly improves the semantic consistency expression of complex components such as elongated, thin-walled, and sequential at the detail level, and improves the problems of boundary blur, morphological distortion, and label discontinuity in the narrowband structure detection of the ship body of the traditional semantic segmentation model, thereby providing a more accurate structure segmentation basis for the anticorrosion spraying quality detection. Through the above perception symmetry loss mechanism, the model not only learns the label space consistency of the image, but also learns the overall symmetry rule of the ship body structure in the deep semantic space. This method significantly improves the robustness of the model to symmetrical structures such as bulkhead symmetrical components, rib plate sequences, and frame partitions, reduces the problems of false classification and structure interruption caused by local texture differences, and especially shows better boundary integrity and label consistency in complex structure areas, thereby providing a reliable structure semantic basis for subsequent fine segmentation of the spraying area and structure-process matching.

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

[0089] In the foregoing continuation example, the internal panoramic image of the ballast tank collected and spliced by Insta360Pro2 is subjected to structure semantic recognition, and the system uses the trained SegFormer-B5 network model as the segmentation backbone network to identify five structure regions of the tank top, longitudinal bulkhead, bottom plate, rib plate, and ladder in the image. Each region is assigned a corresponding semantic label to provide a structure basis for subsequent spraying strategy calling based on position matching. The final output is a structure semantic image of the same size as the input image, in which different colors represent different structure parts, serving as an index basis for local image structure re-localization in the next step.

[0090] S3, according to the structure semantic label, establishing a structure-spraying process mapping rule to form a mapping relationship between the structure type and the corresponding spraying type.

[0091] To realize the intelligent judgment of structure-based spraying process, the identified structure semantic labels need to be mapped with the corresponding anticorrosion spraying process specifications. Due to the differences in the environment, stress state and use function of different structure parts of a ship, the applicable anticorrosion spraying types also have significant differences. For example, the parts below the waterline need to use antifouling paint with seawater immersion and biological attachment performance, while the interior of the ballast tank needs to use thick film epoxy paint with strong adhesion and high pressure resistance. Therefore, by establishing the mapping rules between structure and spraying, the automatic conversion of structure semantic labels to spraying detection parameters can be realized, supporting the spraying type judgment of local images and the subsequent quality detection strategy loading.

[0092] The spraying type refers to the anticorrosion coating system specified for a specific structure part, including the primer type, intermediate coating and topcoat configuration, as well as the target thickness, anticorrosion performance requirements and detection standards of each coating layer.

[0093] The implementation process of this step S3 includes the following operations: first, according to the existing ship classification society standards, engineering specifications and historical spraying process data, a structure-spraying mapping rule library is constructed. The rule library includes information such as structure semantic labels, recommended spraying system, target film thickness range, typical defect types to be detected and qualified judgment threshold. The optional implementation scheme includes: establishing a static dictionary mapping between structure labels and spraying strategies in the form of key-value pairs, or using a decision tree-based logic rule set to realize the reasoning of structure to spraying parameters.

[0094] In specific implementation, the system receives a structure semantic graph as input, calls the mapping rule library for each label 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 structure category (such as "cabin interior"), then refining to the specific part (such as "ballast tank wall"), and further matching the spraying environment (such as "normal temperature construction" or "moisture high corrosion area") to select the optimal spraying system.

[0095] In a preferred implementation, to adapt to the differences in spraying strategies between different ship types and different projects, the structure-spraying mapping rule library can be designed as a configurable JSON structure or a database table structure, supporting manual editing and version management, and allowing the interface to be connected to the shipyard design process management system to realize automatic synchronization.

[0096] If the structure semantic set is S and the spraying type set is T, the structure-spraying mapping relationship can be represented as a mapping function:

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

[0098] Wherein, s represents the structure semantic label, T(s) represents the corresponding spraying type, f represents the decision function of the structure label mapping to the spraying rule, and the specific form can be a lookup table mapping, a conditional reasoning function or a machine learning classifier.

[0099] Continuing the previous example, after the semantic recognition of the bulk carrier ballast tank panoramic image, the area label is recognized as "longitudinal bulkhead", and the system queries the rule base to know that the structure should adopt "thick film epoxy system", wherein the primer is solvent-free epoxy primer, the target thickness is 300 to 350 microns, and the quality detection needs to focus on three defect indicators of blistering, peeling and adhesion abnormality. The spraying type information is used as the spraying parameter input for the subsequent local image detection

[0100] S4, acquiring a local image, using a handheld device, a mobile terminal or a detection robot to acquire a local image of the sprayed area of the ship body.

[0101] In order to realize the detail quality analysis of the sprayed area of the ship body, it is necessary to acquire images of local areas after the spraying operation is completed. Since spraying defects such as blistering, missing spraying and sagging usually occur at the microscale or local range, they cannot be directly recognized by panoramic images, so high-resolution images need to be acquired in key areas to support subsequent defect detection, film thickness estimation and surface state evaluation. Local image acquisition not only provides detailed information, but also serves as an input for matching the structure semantics of the panoramic image, forming an intermediate bridge between structure positioning and spraying detection.

[0102] The local image refers to high-resolution image data acquired in a local area, usually with a resolution better than 0.2 millimeters per pixel, which can clearly reflect surface texture, color uniformity and detail defects. The handheld device refers to an image acquisition device held by a person, such as an industrial camera, a mobile terminal or a structured light scanner. The detection robot refers to a robot platform that can move on the surface of the ship body or autonomously patrol in the cabin, carrying an image sensor to perform image acquisition tasks.

[0103] The specific implementation process of this step S4 is as follows: after the spraying operation is completed, according to the preset detection path or according to the position indicated by the structure semantic map, use the image acquisition device to acquire the local image. The optional implementation scheme includes: a detection personnel holds a high-definition camera or a terminal device equipped with a light ring to take pictures in the target area; or uses a track-type, magnetic-type or aerial flying detection robot equipped with a vision module to complete image acquisition during structure surface inspection. During the acquisition process, in order to ensure image quality, the imaging distance should be controlled within a preset range (such as 30 to 70 centimeters), and the vertical angle should be 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 the situation of insufficient light in the ship body 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 this step S5 is as follows: first, the image features of the local image are extracted, which include but are not limited to color histogram, gradient direction histogram, SIFT local feature point set, ORB key point set, or semantic feature vector output by the intermediate layer of the deep network. Then the feature set is matched with the feature set of each candidate region in the panoramic image to calculate the similarity and find the best matching region. The optional implementation scheme includes: using the brute force sliding window matching method to traverse and match all positions of the panoramic image, or using the KD tree accelerated approximate nearest neighbor algorithm to quickly match the feature vectors. Preferably, local sensitive hashing or multi-scale comparison mechanism based on feature pyramid can be introduced to improve the matching speed and robustness.

[0113] After matching, the system obtains the structure label of the mapping position in the panoramic structure semantic graph, thereby determining which hull structure region the local image belongs to.

[0114] In a preferred implementation, to improve the reverse positioning accuracy of the local image in the panoramic image, the local image acquisition device can also carry a pose perception module, such as an inertial measurement unit (IMU), a two-dimensional code identification scanning system, a visual-inertial integrated positioning module (VIO), or other sensors that can obtain displacement and orientation. The above-mentioned pose perception module can obtain the three-dimensional spatial position information and attitude angle (including pitch angle, roll angle and yaw angle) of the acquisition device at the moment of shooting in real time, thereby providing prior parameter constraints for the spatial matching between the local image and the panoramic image, reducing the search range, reducing the matching ambiguity, and improving the uniqueness and stability of the reverse positioning.

[0115] Specifically, when the local image is acquired, the IMU can record acceleration and angular velocity data synchronously, and after Kalman filtering or sliding window optimization, the three-dimensional trajectory and attitude information of the device can be calculated. This information can be preliminarily aligned with the three-dimensional spherical coordinate system or the projection coordinate system established by the panoramic image, serving as the initial value of the spatial constraint of image matching. On the other hand, the two-dimensional code identification scanning system can quickly anchor the local image to the predefined reference coordinate system or the calibration point by identifying the two-dimensional code icon number and relative position pre-set in the ship body, thereby positioning its approximate area in the large-scale panoramic image, and then realizing accurate mapping through image content detail matching.

[0116] In a specific implementation, the pose data can be input as parameters to the image feature matching module to screen and weight the spatial consistency of the matching point pairs, for example, while calculating the similarity of the image feature descriptor, a pose consistency scoring item is introduced, so that the candidate regions with close viewing angle and spatial position obtain higher matching confidence. In addition, if the system has a time synchronization mechanism, the pose trajectory and the panoramic shooting path can also be jointly optimized to further improve the global consistency of the local image positioning.

[0117] This mechanism is particularly suitable for the internal closed cabin section of the ship body or the area with strong structural repetition. In the condition of lack of texture, large illumination change or high similarity of multi-region images, the false matching rate can be significantly reduced, the uniqueness and stability of the positioning result are ensured, and a reliable position basis is provided for the subsequent spraying process judgment and quality analysis.

[0118] If the feature vector of the local image is F_local, and the feature of the panoramic image at the coordinate position (x, y) is F_pano(x, y), then the 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 measurement function, which can be selected as Euclidean distance, cosine distance or Mahalanobis distance, and (x_hat, y_hat) is the position of the optimal matching position in the panoramic image coordinate system.

[0121] Continuing the previous example, in the detection of the cabin wall image collected by the personnel, the SIFT feature point set is extracted, and the system matches it with the feature library of the ballast tank panoramic image to locate the best matching position at the horizontal angle of 135 degrees and the vertical angle of 40 degrees of the panoramic image, corresponding to the structure label "longitudinal bulkhead".

[0122] S6. According to the structure semantic label, determine the spraying type and spraying quality detection standard corresponding to the local image.

[0123] In order to realize the accurate detection goal of structure attribute, it is necessary to determine the spraying type and spraying quality detection standard corresponding to the local image according to the structure semantic label of the local image in the panoramic image. Since there are significant differences in the types of anticorrosive coatings, process parameters and quality evaluation standards used by each structure part of the ship body, if the corresponding detection rules are not loaded according to the structure attribute, it may lead to false defect judgment or missed detection problem. Therefore, through the parameter matching mechanism driven by the structure semantic label, the detection process of each local image can have process relevance, detection parameter consistency and evaluation standard adaptability.

[0124] The spraying quality detection standard refers to the detection parameter threshold and judgment rule formulated for a certain spraying type, which usually includes the target film thickness range, the allowed color difference value, the bubble area threshold value, the sag length tolerance value, etc., and is configured according to the ship classification society standard or the coating supplier technical specification.

[0125] The specific implementation process of this step S6 is as follows: first, the system receives the structural semantic label obtained in step S5, and calls the corresponding spraying type according to the pre-established structure-spraying mapping relationship library. Subsequently, the system matches the corresponding quality detection standard according to the spraying type, and forms a complete set of detection parameters. The parameter set can include a lower limit value T_min of film thickness, an upper limit value T_max of film thickness, a color difference tolerance value Delta_E_max, a maximum size A_bubble_max of bubbles, an allowable defect pixel rate P_defect_max, etc. The optional implementation scheme includes: using a static dictionary structure to establish a direct mapping of structure to detection parameters, or constructing a process decision engine based on rule expression.

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

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

[0128] If the structure semantic label is S, the spraying type is T, and the detection standard set is Q, then:

[0129] T = f(S)

[0130] Q = g(T)

[0131] Where f represents the mapping function of structure semantic label to spraying type, g represents the query function of spraying type to detection parameter set, S is the structure label input, T is the spraying type output, and Q is the parameter set containing multiple detection indicators.

[0132] Continuing the foregoing example, the system has determined the structure semantic label of the local image to be "longitudinal bulkhead" in step S5, and has found in the mapping rule library that its spraying type is a thick film epoxy system. The system then loads its quality detection standard: the film thickness range is 280 to 350 microns, the allowable color difference value is not greater than 2.0, the bubble diameter is not more than 1 millimeter, and the cumulative area is not more than 0.5% of the image area. This parameter set serves as the input threshold basis for the subsequent local image automatic detection model, ensuring that the evaluation process meets the actual engineering standards.

[0133] S7, spraying quality detection is performed on the local image, and an image recognition-based defect detection model is used to analyze the quality of the spraying area in the local image, and output a quality evaluation result.

[0134] The spraying defect detection model refers to an image recognition algorithm system taking an image as input and outputting a defect region, type and degree determination. It is usually based on a deep learning model and is equipped with pre-defined defect categories and detection parameters.

[0135] The specific implementation process of this step S7 is as follows: first, input the local image into the spraying defect detection model, extract the multi-scale features of the image, and identify the typical spraying defect region in the image. The defect types include but are not limited to missing spraying, sagging, bubbling, peeling, insufficient film thickness, color difference abnormalities, etc.

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

[0137] The system judges the compliance of the identified defect region according to the quality detection standard corresponding to the structure-spraying type. For example, if the detected bubble area is greater than the set threshold value, or the film thickness estimate value is lower than the lower limit value, it is determined that the image is unqualified, and the corresponding position and defect type are recorded. After the detection is completed, the quality evaluation result is output, including whether it is qualified, the type of defect, the defect position coordinates, the defect area ratio, the film thickness range, the color difference value and other indicators.

[0138] In a preferred implementation, the defect detection model can introduce a multi-modal fusion mechanism, inputting the RGB image, infrared thermal image and structured light depth map into the model together, to improve the recognition accuracy of hidden defects such as bubbling and film thickness fluctuation. Further, to improve the detection efficiency and processing speed, a lightweight model such as MobileNet-DeepLab structure can be used and deployed on an edge computing device to support real-time detection on site.

[0139] If the image input is I_local and the model output is D_local, 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 defect information set output by the detection model, Q_standard is the quality judgment parameter set corresponding to the structure-spraying standard, and h represents the evaluation function, which compares the detection result with the standard and outputs whether it is qualified and various quantitative indicators.

[0142] The implementation of this step has the following beneficial effects: realizing the automation and intelligentization of the spraying quality detection process, significantly improving the detection efficiency and coverage, reducing the manual error and the risk of missed detection, having the advantages of structure and semantic association, defect accurate identification, index quantitative output, etc., and providing a reliable foundation for the closed-loop control of the spraying process quality.

[0143] Continuing the previous example, the system has determined that the local image belongs to the "longitudinal bulkhead" structural part, and the corresponding spraying type is the thick film epoxy system. The system inputs the image into the U-Net semantic segmentation model deployed in the edge terminal, and the detection result shows that there is a blister area with an area of 1.2 square centimeters, the film thickness is estimated to be 180 microns, which is lower than the lower limit of the standard 280 microns, and the color difference value is 1.5, which is within the allowable range. The system outputs a detection report that the image is unqualified and marks the defect position for rework processing or subsequent tracking verification.

[0144] In another embodiment, the present application also provides a ship anticorrosion spraying quality detection system based on machine vision, comprising:

[0145] A panoramic image acquisition module is configured to acquire a 360-degree panoramic image of a space to be detected of a ship body, and a panoramic imaging device is configured to perform panoramic shooting on the external and / or internal structure of the ship to form a spherical panoramic image data;

[0146] A structure semantic recognition module is configured to perform structure semantic recognition on the panoramic image, and a semantic segmentation network is configured to divide the panoramic image into a plurality of regions with structure semantic labels, each semantic label corresponding to a specific structural part of the ship body;

[0147] A spraying rule matching module is configured to establish a structure-spraying process mapping rule according to the structure semantic labels, and form a mapping relationship between the structure type and the corresponding spraying type;

[0148] A local image acquisition module is configured to acquire a local image of a sprayed area of the ship body by a handheld device, a mobile terminal or a detection robot;

[0149] A spatial position reverse positioning module is configured 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 structure semantic label;

[0150] A spraying standard extraction module is configured to determine the spraying type and the spraying quality detection standard corresponding to the local image according to the structure semantic label;

[0151] A quality detection module is configured to perform spraying quality detection on the local image, and a defect detection model based on image recognition is configured to perform quality analysis on the spraying area in the local image, and output a quality evaluation result.

[0152] Preferably, the structural semantic recognition module is configured to employ a semantic segmentation network introducing an auxiliary loss function during training, the auxiliary loss function being constructed based on structural symmetry and topological constraints, and a total loss function being:

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

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

[0155] Preferably, the structural semantic recognition module further comprises a symmetric structure perception module for introducing a symmetric perception loss to construct a structural semantic symmetry constraint, the symmetric perception loss being constructed in the following manner:

[0156] A mirror image I_mirror is generated based on each prediction result I_pred, and features F_l(I_pred) and F_l(I_mirror) at the lth layer of a pre-trained classification network are extracted, and a perception symmetry loss function is constructed:

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

[0158] wherein w_l is a feature layer weighting coefficient, F_l is the feature map at the lth layer, and ||·||2 is an L2 norm;

[0159] The structural semantic recognition module is further configured to weight and fuse the loss function and a basic semantic segmentation loss L_main to construct a total loss function:

[0160] L_total = L_main + λ x L_sym_perceptual,

[0161] wherein λ is a weight adjustment coefficient.

[0162] Preferably, the local image acquisition module comprises a pose perception module working synchronously therewith, the pose perception module being one of an IMU (Inertial Measurement Unit), a two-dimensional code recognition system, or a visual-inertial integrated positioning module, or a combination thereof,

[0163] The pose perception module is used to record acceleration and angular velocity data synchronously when the local image is acquired, and to calculate three-dimensional trajectory and attitude information of the device through Kalman filtering or sliding window optimization, as initial values for spatial repositioning.

[0164] Preferably, the spatial position inverse positioning module is configured to realize position matching by the following way:

[0165] 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) thereof, which satisfies the following relationship:

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

[0167] Wherein, D is a feature distance measurement function, and the feature distance measurement function is one of Euclidean distance, cosine distance or Mahalanobis distance.

[0168] It should be noted that the above-mentioned explanation and description of the ship anti-corrosion spraying quality detection method based on machine vision is also applicable to the device of the embodiment of the present application, and will not be repeated here.

[0169] Those skilled in the art can realize that the units and algorithm steps described in the embodiments disclosed herein can be realized by electronic hardware, computer software and combination of electronic hardware and computer software. Whether the functions are realized in hardware or software mode depends on the specific application and design constraints of the technical solutions. The skilled person can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0170] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-mentioned system, device and unit can refer to the corresponding process in the foregoing method embodiments, which will not be repeated here.

[0171] In several embodiments provided in the present application, any function realized in the form of a software function unit and sold or used as an independent product can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or the part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the embodiments of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (Read-Only Memory; hereinafter referred to as: ROM), a random access memory (Random Access Memory; hereinafter referred to as: RAM), a magnetic disk or an optical disk, and various storage medium capable of storing program codes.

[0172] The above description is only specific embodiments of the present application, and any skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered by the protection scope of the present application. The module structure of the part of the present application not specifically described shall be subject to the content recorded in the prior art. The prior art mentioned in the foregoing background section and the specific embodiment section of the present application can be used as a part of the present application to understand the meaning of some technical features or parameters.

Claims

1. A method for detecting the quality of anticorrosive spraying of a ship based on machine vision, characterized in that, The method comprises the following steps: S1, acquiring a 360-degree panoramic image of a hull space to be detected, performing panoramic shooting on the external and / or internal structure of the ship through a panoramic imaging device to form a panoramic image data in spherical development; S2, performing structural semantic recognition on the panoramic image, dividing the panoramic image into a plurality of regions with structural semantic labels based on a semantic segmentation network, each semantic label corresponding to a specific structural part of the hull; S3, establishing a structure-spraying process mapping rule according to the structural semantic labels to form a mapping relationship between the structure type and the corresponding spraying type; S4, acquiring a local image, collecting a local image of a hull spraying completed area by using a handheld device, a mobile terminal or a detection robot; S5, performing spatial position reverse positioning on the local image, matching the feature information of the local image with the panoramic image, determining the mapping position of the local image in the panoramic image, and acquiring the corresponding structural semantic label; S6, determining the spraying type and spraying quality detection standard corresponding to the local image according to the structural semantic label; S7, performing spraying quality detection on the local image, performing quality analysis on the spraying area in the local image by using a defect detection model based on image recognition, and outputting a quality evaluation result.

2. The machine vision-based marine anti-corrosion spray quality detection method 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 machine vision-based marine anti-corrosion spray quality detection method according to claim 1, characterized in that, The semantic segmentation network uses a symmetric structure-aware perception loss during training, and the symmetric structure-aware perception loss is specifically: For each input image prediction result I_pred, a mirror image I_mirror is obtained according to a predefined structural symmetry axis; An image classification network pre-trained on a general dataset is used as a perception feature extractor to extract the output of an intermediate feature layer respectively; An image classification network pre-trained on a general dataset is used as a perception feature extractor to extract the l-th layer features F_l(I_pred) and F_l(I_mirror) of the image prediction result I_pred and the mirror image I_mirror, and the perception symmetry loss is defined as the Euclidean distance between the l-th layer features. The perception symmetry loss and the main segmentation loss function are combined to form a total loss function.

4. The machine vision-based marine anti-corrosion spray quality detection method according to claim 1, characterized in that: When the local image is collected, the IMU synchronously records acceleration and angular velocity data, and after Kalman filtering or sliding window optimization, the three-dimensional trajectory and attitude information of the device are calculated, and then the three-dimensional spherical coordinate system or the projection coordinate system established with the panoramic image is preliminarily aligned as the spatial constraint initial value of image matching.

5. The machine vision-based marine anti-corrosion spray quality detection method according to claim 1, characterized in that The spatial position reverse positioning on the local image comprises: 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).

6. A machine vision-based ship anticorrosion spray quality detection system, characterized by, The system comprises the following modules: The panoramic image acquisition module is configured to acquire a 360-degree panoramic image of a space to be detected of a ship body, and to form spherical panoramic image data by panoramic imaging of an external and / or internal structure of the ship body. The structural semantic recognition module is configured to perform structural semantic recognition on the panoramic image, and to divide the panoramic image into a plurality of regions with structural semantic labels based on a semantic segmentation network, each semantic label corresponding to a specific structural part of the ship body. The spraying rule matching module is configured to establish a structure-spraying process mapping rule according to the structural semantic labels, and to form a mapping relationship between a structural type and a corresponding spraying type. The local image acquisition module is configured to acquire a local image of a sprayed region of the ship body by using a handheld device, a mobile terminal or a detection robot. The spatial position back positioning module is configured to match feature information of the local image with the panoramic image, to determine a mapping position of the local image in the panoramic image, and to obtain a corresponding structural semantic label. The spraying standard extraction module is configured to determine a spraying type and a spraying quality detection standard corresponding to the local image according to the structural semantic label. The quality detection module is configured to perform spraying quality detection on the local image, to analyze a spraying region in the local image by using a defect detection model based on image recognition, and to output a quality evaluation result.

7. The machine vision-based marine anti-corrosion spray quality inspection system according to claim 6, wherein, The structural semantic recognition module is configured to use a semantic segmentation network with an auxiliary loss function introduced during training, and the auxiliary loss function is constructed based on structural symmetry and topological constraints.

8. The machine vision-based marine anti-corrosion spraying quality detection system according to claim 6, wherein, The structural semantic recognition module further includes a symmetric structure perception module configured to introduce a symmetric perception loss to construct a structural semantic symmetry constraint, and the symmetric perception loss is constructed in the following manner: A corresponding mirror image I mirror is generated based on each prediction result I_pred, and features F_l(I_pred) and F_l(I mirror) at the lth layer of a pre-trained classification network are extracted, and a symmetric perception loss function is constructed. The structural semantic recognition module is further configured to weight and fuse the loss function and a basic semantic segmentation loss L_main to form a total loss function.

9. The machine vision-based marine anti-corrosion spray quality inspection system according to claim 6, wherein, The local image acquisition module includes a pose perception module that works synchronously therewith, the pose perception module being one of an IMU (inertial measurement unit), a two-dimensional code recognition system or a visual-inertial combined positioning module, or a combination thereof. The pose perception module is configured to record acceleration and angular velocity data synchronously when the local image is acquired, and to calculate three-dimensional trajectory and attitude information of the device by Kalman filtering or sliding window optimization, as initial values for spatial back positioning.

10. The machine vision-based marine anti-corrosion spray quality inspection system according to claim 6, wherein, The spatial position back positioning module is configured to realize position matching in the following manner: A feature vector F_local of the local image is matched with a feature vector F_pano(x,y) of a coordinate position (x,y) of the panoramic image to determine a matching position (x_hat,y_hat).

Images