Artificial Intelligence-Based Method and System for Detecting Appearance Defects in Concrete of Underwater Structures

Abstract

This invention discloses an artificial intelligence-based method and system for detecting surface defects in underwater concrete structures. The method includes: firstly, calibrating preset color and monochrome cameras both aerially and underwater. Then, acquiring first and second images (with and without laser stripes) taken by the calibrated color camera, and a third image taken by the monochrome camera. Based on matching the laser stripes in the first and third images, a point cloud image is constructed and combined with color enhancement of the second image to obtain a target point cloud image. Finally, a surface defect detection model is used to identify defects in the image, yielding the defect detection result, thus achieving efficient and accurate detection of surface defects in underwater concrete structures.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to a method and system for detecting apparent defects in concrete of underwater structures based on artificial intelligence. Background Technology

[0002] With the increasing number of underwater structures, their safety has become increasingly important. Detecting surface defects in the concrete of underwater structures faces many challenges, as traditional detection methods suffer from low accuracy, poor efficiency, and are greatly affected by the underwater environment. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for detecting apparent defects in concrete of underwater structures based on artificial intelligence.

[0004] In a first aspect, embodiments of the present invention provide a method for detecting apparent defects in concrete of underwater structures based on artificial intelligence, comprising:

[0005] Perform aerial and underwater calibration on the preset color camera and the preset monochrome camera;

[0006] Acquire a first image and a second image of the concrete appearance of an underwater structure taken by a calibrated color camera. The first image has laser stripes, and the second image does not have laser stripes. The laser stripes are generated based on underwater refraction.

[0007] Acquire a third image of the concrete appearance of an underwater structure taken by a calibrated black-and-white camera;

[0008] The first image and the third image are matched based on the laser stripes. If the match is successful, a viewpoint cloud image of the concrete surface of the underwater structure is constructed based on the first image and the third image. Color enhancement is performed based on the second image to obtain the viewpoint cloud image of the concrete surface of the target underwater structure.

[0009] Based on the apparent defect detection model, defect identification is performed on the surface cloud image of the concrete of the target underwater structure to obtain the defect detection result corresponding to the surface cloud image of the concrete of the target underwater structure.

[0010] In one possible implementation, the defect identification of the target underwater structure's concrete surface cloud image based on the apparent defect detection model, to obtain the defect detection result corresponding to the target underwater structure's concrete surface cloud image, includes:

[0011] A first sample building concrete appearance image set and a second sample building concrete appearance image set are obtained. The first sample building concrete appearance image set includes first sample building concrete appearance images, and the second sample building concrete appearance image set includes a second sample building concrete appearance image set configured with sample target values. The second sample building concrete appearance image set is classified into a target defect category set.

[0012] The main defect detection model is obtained by training the first sample building concrete appearance image set with a multi-task learning method and the second sample building concrete appearance image set with a reinforcement learning method.

[0013] Obtain the defect detection model, wherein the defect detection model is the original model whose model parameters have not been trained;

[0014] Using the first sample building concrete appearance image as the teacher model and the main defect detection model as the teacher model, knowledge transfer is performed on the model parameters of the secondary defect detection model to obtain the appearance defect detection model;

[0015] The apparent defect detection model is used to identify defects in the surface cloud image of the concrete of the target underwater structure, and the defect detection result of the surface cloud image of the concrete of the target underwater structure in the target defect category set is obtained.

[0016] In one possible implementation, the step of using the first sample building concrete appearance image as a teacher model and performing knowledge transfer on the model parameters of the defect detection model to obtain the appearance defect detection model includes:

[0017] The main defect detection model is used to identify defects in the concrete appearance images of the first sample building in the first sample building concrete appearance image set, and to obtain the undetermined target value of the category in the target defect category set corresponding to the concrete appearance image of the first sample building.

[0018] By performing defect identification on the concrete appearance images of the first sample building in the first sample building concrete appearance image set using the defect detection model, the identification result corresponding to the concrete appearance image of the first sample building is obtained.

[0019] The model parameters of the defect detection model are optimized based on the deviation between the target value to be determined and the identification result to obtain the apparent defect detection model.

[0020] In one possible implementation, optimizing the model parameters of the defect detection model based on the deviation between the target value to be determined and the identification result to obtain the apparent defect detection model includes:

[0021] Obtain the credibility index of the target value to be determined;

[0022] The deviation between the target value to be determined and the identification result under the credibility index is determined, and the model parameters of the defect detection model are optimized based on the deviation to obtain the apparent defect detection model.

[0023] In one possible implementation, the step of identifying defects in the concrete appearance images of the first sample building in the first sample building concrete appearance image set using the master defect detection model, and obtaining the undetermined target value of a category in the target defect category set corresponding to the concrete appearance image of the first sample building, includes:

[0024] The main defect detection model is used to identify defects in the concrete appearance images of the first sample building in the first sample building concrete appearance image set, and the confidence level of the first sample building concrete appearance image in the target defect category set is obtained.

[0025] Based on the confidence level, the target value corresponding to the concrete appearance image of the first sample building is determined from the target defect category set.

[0026] In one possible implementation, the defect detection model includes a first index vectorization unit and a first related element vectorization unit;

[0027] The method further includes:

[0028] The first sample building concrete appearance image is vectorized by the first index vectorization unit to obtain a first vector representation;

[0029] The first relevant element vectorization unit obtains a second vector representation of the first sample building concrete appearance image and the data in the first preset database;

[0030] Based on the deviation between the first vector representation and the second vector representation, knowledge transfer is performed on the defect detection model to obtain the apparent defect detection model.

[0031] In one possible implementation, the step of training the main defect detection model using a multi-task learning method with the first set of sample building concrete appearance images and a reinforcement learning method with the second set of sample building concrete appearance images includes:

[0032] Multi-task learning is performed on the feature extraction component based on the concrete appearance images of the first sample building in the first sample building concrete appearance image set.

[0033] By combining the gated recurrent component with the feature extraction component that has undergone multi-task learning, a defect detection model is obtained. The gated recurrent component is used to perform appearance defect detection in the target defect category set.

[0034] The defect detection model is obtained by performing reinforcement learning on the defect detection model using the second sample building concrete appearance image set and the sample target value.

[0035] In one possible implementation, the feature extraction component includes a second index vectorization unit and a second related element vectorization unit;

[0036] The feature extraction component, based on the first sample building concrete appearance images in the first sample building concrete appearance image set, performs multi-task learning, including:

[0037] The first sample building concrete appearance image is vectorized by the second index vectorization unit to obtain a third vector representation;

[0038] The second related element vectorization unit obtains a fourth vector representation of the first sample building concrete appearance image and the data in the second preset database;

[0039] The feature extraction component performs multi-task learning based on the deviation between the third vector representation and the fourth vector representation.

[0040] In one possible implementation, combining the gated recurrent component with the feature extraction component learned through multi-task training to obtain the defect detection model includes:

[0041] The gated loop component is concatenated with the second index vectorization unit, which has undergone multi-task learning, to obtain the defect detection model.

[0042] In a second aspect, embodiments of the present invention provide a server system, including a server, the server being used to execute the method described in the first aspect.

[0043] Compared to existing technologies, the beneficial effects provided by this invention include: The method and system for detecting apparent defects in underwater concrete structures based on artificial intelligence, as disclosed in this invention, involves calibrating preset color and monochrome cameras both aerially and underwater. Then, first and second images with and without laser stripes, captured by the calibrated color camera, and a third image captured by the monochrome camera are acquired. Based on matching the laser stripes in the first and third images, a point cloud image is constructed and combined with color enhancement of the second image to obtain the target point cloud image. Finally, an apparent defect detection model is used to identify defects in the image, yielding the defect detection result, thus achieving efficient and accurate detection of apparent defects in underwater concrete structures. Attached Figure Description

[0044] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as limiting the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 A flowchart illustrating the steps of an artificial intelligence-based method for detecting apparent defects in concrete structures in underwater structures, as provided in an embodiment of the present invention.

[0046] Figure 2 A schematic block diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0048] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0049] In order to solve the technical problems mentioned in the background art Figure 1 This is a flowchart illustrating the artificial intelligence-based method for detecting apparent defects in concrete of underwater structures, as provided in this embodiment. The following is a detailed description of this artificial intelligence-based method for detecting apparent defects in concrete of underwater structures.

[0050] Step S201: Perform aerial calibration and underwater calibration on the preset color camera and the preset black and white camera;

[0051] Step S202: Acquire a first image and a second image of the concrete appearance of an underwater structure taken by a calibrated color camera. The first image has laser stripes, and the second image does not have laser stripes. The laser stripes are generated based on underwater refraction.

[0052] Step S203: Acquire a third image of the concrete appearance of an underwater structure taken by a calibrated black and white camera;

[0053] Step S204: Match the first image and the third image based on the laser stripes. If the matching is successful, construct a view cloud image of the concrete surface of the underwater structure based on the first image and the third image, and enhance the color based on the second image to obtain the view cloud image of the concrete surface of the target underwater structure.

[0054] Step S205: Based on the apparent defect detection model, perform defect identification on the surface cloud image of the concrete of the target underwater structure to obtain the defect detection result corresponding to the surface cloud image of the concrete of the target underwater structure.

[0055] In an embodiment of the present invention, for example, at the construction site of a large-scale water conservancy project, the server receives a task to detect defects in the concrete surface of underwater structures. First, the preset color camera and the preset black-and-white camera need to be calibrated.

[0056] Aerial calibration: A color camera and a monochrome camera are mounted on a precisely controllable robotic arm, placed in an open space with clearly defined features, such as a room filled with regular patterns. The server controls the robotic arm to move the cameras along a pre-set trajectory, while the cameras capture images from different positions and angles. These images contain rich spatial feature information. By analyzing the positions of feature points in these images and the camera's movement parameters, the server calculates the camera's internal parameters (such as focal length and principal point position) and external parameters (such as the camera's position and orientation in space). This is the process of aerial calibration, which lays the foundation for accurate shooting and image analysis in underwater environments.

[0057] Underwater calibration: The color and monochrome cameras, already calibrated in the air, are mounted in a specially designed waterproof enclosure and placed in a large water tank. Inside the tank are standard objects with known dimensions and positions, such as specially designed standard geometric modules. The server controls the cameras to capture images of these standard objects underwater. Due to the refraction and absorption properties of water, the images will exhibit some distortion. The server analyzes these distorted images and, combined with the known size and position information of the standard objects, further corrects and optimizes the camera parameters to adapt to the special optical conditions of the underwater environment. This process is called underwater calibration, which ensures that the images captured by the camera underwater accurately reflect the true shape and position of the objects.

[0058] After camera calibration, the color camera is mounted on the underwater robot, and the underwater robot is placed below the dam of the hydraulic engineering project. The server controls the underwater robot to approach the concrete surface of the underwater structure.

[0059] When the color camera begins shooting, it first captures an initial image with laser stripes. The server controls a laser emitter mounted on the underwater robot to emit laser beams of a specific pattern. When these beams contact the concrete surface of the underwater structure, the refraction of the water creates unique laser stripes. The color camera captures these images with laser stripes and transmits them back to the server.

[0060] Subsequently, the server controls a color camera to capture a second image without laser stripes. This is to obtain image information of the concrete surface under normal lighting conditions for subsequent color enhancement and comprehensive analysis.

[0061] While the color camera is capturing images, a monochrome camera mounted on the underwater robot is also working synchronously. The monochrome camera captures high-contrast images of the concrete surface of underwater structures, demonstrating excellent ability to capture the contours and details of objects. The server acquires the third set of images from the monochrome camera; these images complement the images from the color camera, providing more comprehensive information for subsequent point cloud image construction.

[0062] After receiving the first and third images, the server begins image matching. Because the laser stripes have a unique distribution and characteristics in the first image, the server first extracts feature points from the laser stripes in the first image, such as the intersections and endpoints of the stripes. Then, it searches for corresponding feature points in the third image. Through feature point matching and calculation, the correspondence between the first and third images is determined.

[0063] After a successful match, the server constructs a point cloud image of the underwater structure's concrete surface based on the first and third images. The server calculates the 3D coordinates of each point on the object's surface according to the camera's calibration parameters and the location of feature points in the images. By connecting a large number of points, a point cloud image reflecting the shape of the underwater structure's concrete surface is formed.

[0064] Simultaneously, the server acquires the second image and performs color enhancement processing on it. The server analyzes the color distribution and brightness values ​​of the pixels in the second image, and enhances the vibrancy and clarity of the colors in the image by adjusting parameters such as contrast, saturation, and brightness, making the details of the concrete surface more apparent.

[0065] Finally, the color-enhanced second image is fused with the constructed point cloud image to obtain a point cloud image of the concrete surface of the target underwater structure. This fused image contains both precise three-dimensional shape information of the object's surface and rich, clear color information, providing high-quality input data for subsequent defect detection.

[0066] The server has obtained a cloud image of the concrete surface of the target underwater structure. The next step is to use a pre-trained appearance defect detection model to identify defects.

[0067] Before training the apparent defect detection model, the server first collected a large number of sample building concrete appearance images, including images that were normal and those with various defects. These images were divided into a first sample building concrete appearance image set and a second sample building concrete appearance image set. The first sample building concrete appearance image set was used for multi-task learning to train the model's basic feature extraction capabilities; the second sample building concrete appearance image set, with sample target values, was used for reinforcement learning to optimize the model's defect detection capabilities.

[0068] The server trains a feature extraction component using a multi-task learning approach with a set of initial sample images of concrete surfaces from buildings. For example, given a set of concrete surface images under different lighting conditions, the server teaches the model how to extract common features, such as texture and shape, while also learning how to adapt to different lighting variations.

[0069] The server then combines the gated recurrent component with the feature extraction component, which has undergone multi-task learning, to form a preliminary defect detection model. The gated recurrent component is capable of effectively detecting apparent defects within the target defect category set.

[0070] Next, the server uses a second set of concrete surface images of the building as a sample and the corresponding sample target values ​​to perform reinforcement learning on the defect detection model. For example, given a set of concrete surface images known to have defects such as cracks and spalling, the server adjusts the model's parameters to enable it to accurately identify these defects and provide corresponding classifications and evaluations.

[0071] After repeated training and optimization, the server obtained the primary defect detection model. Then, the server acquired the secondary defect detection model, which was a raw model with untrained parameters. The server used the primary defect detection model to identify defects in images from the first set of sample building concrete appearance images, obtaining the target values ​​to be determined. For example, for a concrete surface image, the primary defect detection model might give an 80% confidence level that it belongs to a crack defect.

[0072] Simultaneously, the server instructs the secondary defect detection model to perform defect identification on the same image, obtaining the identification results. The server optimizes the secondary defect detection model's parameters based on the deviation between the target value and the identification result. For example, if the secondary defect detection model identifies a spalling defect, which deviates significantly from the target value (crack defect) given by the primary defect detection model, the server will adjust the secondary defect detection model's parameters to more closely approximate the primary defect detection model's judgment.

[0073] The server also obtains a confidence index for the target value to be determined. For example, if the main defect detection model has a high confidence level in judging a certain image, it means that the confidence level of its target value to be determined is high; conversely, if the confidence level is low, the confidence level is low. The server optimizes the model parameters of the defect detection model more accurately based on the deviation under the confidence index, and finally obtains the appearance defect detection model.

[0074] Finally, the server inputs the visual cloud image of the target underwater structure's concrete surface into the appearance defect detection model. The model analyzes and judges each area in the image, identifying the presence of defects, as well as the type, location, and severity of the defects. The server obtains the defect detection results corresponding to the visual cloud image of the target underwater structure's concrete surface, such as detecting cracks of a certain length and width at a certain location, or minor spalling at another location. These results will be output and provided to relevant engineers so that they can take appropriate repair and reinforcement measures to ensure the safety and stability of the underwater structure.

[0075] In this embodiment of the invention, the defect identification of the concrete surface cloud image of the target underwater structure based on the apparent defect detection model, and the obtaining of the defect detection result corresponding to the concrete surface cloud image of the target underwater structure, can be implemented through the following example.

[0076] A first sample building concrete appearance image set and a second sample building concrete appearance image set are obtained. The first sample building concrete appearance image set includes first sample building concrete appearance images, and the second sample building concrete appearance image set includes a second sample building concrete appearance image set configured with sample target values. The second sample building concrete appearance image set is classified into a target defect category set.

[0077] The main defect detection model is obtained by training the first sample building concrete appearance image set with a multi-task learning method and the second sample building concrete appearance image set with a reinforcement learning method.

[0078] Obtain the defect detection model, wherein the defect detection model is the original model whose model parameters have not been trained;

[0079] Using the first sample building concrete appearance image as the teacher model and the main defect detection model as the teacher model, knowledge transfer is performed on the model parameters of the secondary defect detection model to obtain the appearance defect detection model;

[0080] The apparent defect detection model is used to identify defects in the surface cloud image of the concrete of the target underwater structure, and the defect detection result of the surface cloud image of the concrete of the target underwater structure in the target defect category set is obtained.

[0081] In this embodiment of the invention, for example, the server collects a large number of concrete appearance images of buildings from a database of multiple large-scale construction projects. These images cover buildings of different types, with different service lives, and under different environmental conditions.

[0082] The first sample set of building concrete appearance images includes images of concrete surfaces in various normal states as well as those with minor defects. For example, there are images of newly built concrete walls with a good appearance, as well as images of concrete floor slabs that have been in use for several years and have developed some minor cracks that do not affect structural safety.

[0083] The second set of images of concrete structures is a more targeted and severe defect image set, with precise sample target values ​​assigned to each image. These images are categorized into target defect categories. For example, the sample target values ​​for an image of a concrete dam with severe cracks causing leakage clearly indicate detailed information such as the length, width, and depth of the cracks; and the sample target values ​​for an image of a concrete bridge pier with large areas of spalling due to erosion describe in detail the area and depth of the spalled region.

[0084] The server first uses images of the concrete surface of a first sample building for multi-task learning training. For these images, the server is set to perform multiple tasks, such as simultaneously learning to identify different types of defects such as cracks, holes, and corrosion, and adapting to different lighting conditions and shooting angles.

[0085] During training, the server continuously adjusts the model's parameters to extract general features from the image that are applicable to multiple tasks simultaneously. For example, given an image of a concrete wall with both fine cracks and slight corrosion, the model must learn to identify the direction and width of the cracks, as well as to distinguish the color and texture changes in the corroded areas.

[0086] Next, the server uses a second set of images showing the concrete appearance of buildings for reinforcement learning training. In this stage, the server provides explicit reward or penalty signals to the model based on the sample target value of the image.

[0087] For example, given an image of a concrete bridge marked as having severe cracks, if the model can accurately detect key information such as the location, length, and width of the cracks, the server rewards the model significantly; conversely, if the model's detection results differ greatly from the target sample value, the server penalizes the model accordingly. By continuously receiving these feedback signals, the model gradually optimizes its parameters, improving the accuracy and precision of detecting severe defects.

[0088] Through repeated multi-task learning and reinforcement learning training, the server finally obtained a main defect detection model that can accurately detect various defects.

[0089] The server retrieves a brand new, untrained defect detection model from the model library. This model has the same architecture and parameter settings as the main defect detection model, but all parameters are in their initial state and have not been adjusted or optimized with any data.

[0090] The server inputs the first sample building concrete appearance image into the main defect detection model and obtains the output results of the main defect detection model for these images.

[0091] For each input image, the primary defect detection model provides predictions about the possible types and features of defects. The server then passes these predictions as "knowledge" to the secondary defect detection models.

[0092] After receiving this "knowledge" from the defect detection model, it adjusts its own model parameters based on the differences between the output results and those of the main defect detection model.

[0093] For example, if the primary defect detection model predicts the presence of a minor crack in a concrete slab image and provides the approximate location and feature description of the crack, while the initial prediction from the secondary defect detection model differs from this, the server will adjust the relevant parameters of the secondary defect detection model based on this difference, gradually bringing it closer to the prediction of the primary defect detection model.

[0094] Through multiple knowledge transfers and parameter adjustments, the knowledge and experience gained from the defect detection model were gradually learned from the main defect detection model, thus forming the apparent defect detection model.

[0095] After receiving the cloud image of the concrete surface of the target underwater structure, the server inputs it into the appearance defect detection model.

[0096] The model performs comprehensive analysis and processing of the image, extracts various features from the image, and compares and judges them with previously learned defect features.

[0097] Ultimately, the model provides detection results, clearly indicating the specific category to which the image belongs within the target defect category set. For example, the detection results might indicate that the concrete surface of the underwater structure has cracks exceeding a certain threshold in length, or that there is extensive peeling.

[0098] These test results will be recorded and output by the server, providing important information for subsequent repair and maintenance work.

[0099] In this embodiment of the invention, the step of using the first sample building concrete appearance image as the teacher model and performing knowledge transfer on the model parameters of the defect detection model to obtain the appearance defect detection model can be implemented through the following example.

[0100] The main defect detection model is used to identify defects in the concrete appearance images of the first sample building in the first sample building concrete appearance image set, and to obtain the undetermined target value of the category in the target defect category set corresponding to the concrete appearance image of the first sample building.

[0101] By performing defect identification on the concrete appearance images of the first sample building in the first sample building concrete appearance image set using the defect detection model, the identification result corresponding to the concrete appearance image of the first sample building is obtained.

[0102] The model parameters of the defect detection model are optimized based on the deviation between the target value to be determined and the identification result to obtain the apparent defect detection model.

[0103] In an embodiment of the invention, for example, the server acquires a large number of images from a set of first sample building concrete appearance images. One of the images is the concrete surface of a sluice gate that has been in use for a certain number of years, with some minor cracks and localized erosion on the surface.

[0104] The server inputs this image into the master defect detection model. The master defect detection model then performs complex calculations and analysis to identify defects in the image. It comprehensively considers features such as the crack's shape, length, width, and orientation, as well as the area, depth, and texture of the eroded region.

[0105] After a series of calculations and comparisons, the main defect detection model concluded that the defect in this image may belong to the "minor cracks and moderate erosion" category in the target defect category set, and assigned a high confidence level to this judgment, such as 85%. This judgment result became the target value to be determined for this first sample image of the building's concrete appearance.

[0106] The server continues to process other images in the first sample building concrete appearance image set in the same way to obtain the target value to be determined for each image.

[0107] Let's take the image of the concrete surface of the sluice gate from before as an example. The server simultaneously inputs this image into the defect detection model.

[0108] Because the defect detection model has not yet undergone sufficient training and optimization, its identification of defects in images may not be accurate enough. It might conclude that the defects in this image belong to the category of "minor cracks and minor erosion," which differs from the judgment of the main defect detection model.

[0109] For other images in the first sample set of building concrete appearance images, the defect detection model will also provide corresponding recognition results, but these results may deviate to varying degrees from the target values ​​to be determined given by the main defect detection model.

[0110] The server compares the target value to be determined by the main defect detection model with the identification result from the defect detection model. Taking the image of the concrete surface of a sluice gate as an example, there is a discrepancy between the defect category and severity determined by the defect detection model and the target value to be determined by the main defect detection model.

[0111] The server calculates the specific value of this deviation. For example, the deviation is level 1 in crack classification (assuming crack classification is divided into three levels: slight, moderate, and severe), and the deviation is level 1 in erosion degree classification (assuming erosion degree is divided into three levels: mild, moderate, and severe).

[0112] Based on these biases, the server uses specific optimization algorithms to adjust the model parameters from the defect detection model. For example, it may increase the weights associated with crack feature extraction or adjust the thresholds associated with erosion degree assessment.

[0113] The server performs this processing and optimization on each image in the first sample set of concrete appearance images of the building, so that the parameters of the defect detection model are continuously adjusted and improved.

[0114] After multiple iterations and optimizations, the identification results of the defect detection model gradually approached the target value given by the main defect detection model, ultimately yielding the appearance defect detection model. This appearance defect detection model demonstrated higher accuracy and reliability in identifying defects in the concrete appearance images of the first sample building.

[0115] In this embodiment of the invention, the optimization of the model parameters of the defect detection model based on the deviation between the target value to be determined and the identification result to obtain the apparent defect detection model can be implemented through the following example.

[0116] Obtain the credibility index of the target value to be determined;

[0117] The deviation between the target value to be determined and the identification result under the credibility index is determined, and the model parameters of the defect detection model are optimized based on the deviation to obtain the apparent defect detection model.

[0118] In an embodiment of the present invention, for example, when the server processes images from the first sample building concrete appearance image set, it takes an image of the concrete surface of an old bridge pier as an example. After the main defect detection model identifies defects in this image, the resulting target value is "moderate cracks and slight spalling".

[0119] The server obtains a credibility metric for this target value. This credibility metric may be calculated based on multiple factors. For example, a high historical accuracy rate in identifying this type of defect by the main defect detection model, or good image clarity and integrity, allowing the model to extract features more accurately for judgment, can all increase the credibility of the target value. In this example, based on a comprehensive evaluation, the credibility metric for the target value is determined to be 90%.

[0120] The server continues to process other images, such as an image of a dam's concrete surface. The target value given by the main defect detection model is "severe erosion and fine cracks," and its confidence index may be determined to be 70% due to factors such as the complexity of the dam environment and the limitations of the image shooting angle.

[0121] Let's take the images of the old bridge piers and dams mentioned above as examples.

[0122] For images of old bridge piers, the defect detection model identifies "minor cracks and no spalling," which deviates from the target value of "moderate cracks and minor spalling" given by the main defect detection model.

[0123] The server considers a 90% high-confidence metric for the target value and gives greater weight to the assessment of deviations. Specific deviation calculations may involve quantitative comparisons of crack severity and spalling. For example, if crack severity is divided into 1 to 5 levels, with moderate cracks at level 3 and minor cracks at level 2, then the deviation in crack severity is level 1. Differences in the presence or absence of spalling are considered a larger deviation.

[0124] Based on these biases, the server adjusts the relevant model parameters of the defect detection model. For example, it increases the weights associated with crack features or adjusts the thresholds associated with spalling judgment.

[0125] For the dam image, the defect detection model identified "moderate erosion and no cracks," which deviates from the target value of "severe erosion and minor cracks" from the main defect detection model.

[0126] Because the confidence level is 70%, the server will be relatively cautious when assessing bias. However, it will still perform quantitative calculations based on the specific degree of erosion and the presence or absence of cracks, and adjust the parameters of the defect detection model accordingly.

[0127] The server performs this processing on a large number of images from the first sample set of concrete appearance images of buildings. The deviation between the target value to be determined for each image and the recognition result from the defect detection model, taking into account the credibility index, is used to continuously optimize the model parameters from the defect detection model.

[0128] Through multiple iterations and optimizations, the defect detection model gradually learns and adjusts, making its identification results increasingly closer to the target value given by the main defect detection model. Ultimately, a surface defect detection model capable of accurately identifying defects is obtained.

[0129] In this embodiment of the invention, the step of identifying defects in the concrete appearance images of the first sample building in the first sample building concrete appearance image set through the main defect detection model, and obtaining the undetermined target value of the category in the target defect category set corresponding to the concrete appearance image of the first sample building, can be implemented through the following example.

[0130] The main defect detection model is used to identify defects in the concrete appearance images of the first sample building in the first sample building concrete appearance image set, and the confidence level of the first sample building concrete appearance image in the target defect category set is obtained.

[0131] Based on the confidence level, the target value corresponding to the concrete appearance image of the first sample building is determined from the target defect category set.

[0132] In an embodiment of the invention, for example, the server acquires an image from a set of first sample building concrete appearance images, which is an image of the surface of a concrete pillar of a seaside pier. Some subtle texture variations and minor signs of cracks can be seen in the image.

[0133] The main defect detection model begins analyzing the image. It first extracts various features from the image, such as texture direction, crack shape and length, and color variations on the concrete surface. Then, it compares and calculates these features with feature patterns stored internally by the model for various defect types.

[0134] After a series of complex calculations, the main defect detection model arrives at the possible classification of the target defect category to which the image belongs, along with its corresponding confidence level. For example, for this image of a dock support pillar, the main defect detection model determines that it may belong to the "minor crack" category with a confidence level of 75%; it may also belong to the "surface weathering" category with a confidence level of 20%; and there is a smaller possibility that it belongs to the "localized peeling" category with a confidence level of 5%.

[0135] The server continues processing other images in the first set of building concrete appearance images. For example, for an image of the concrete surface of a bridge abutment, the main defect detection model gives a confidence level of 80% for the "moderate corrosion" category, 15% for the "fine cracks" category, and lower confidence levels for other categories.

[0136] For the aforementioned image of the seaside pier pillars, since the "minor crack" category has the highest confidence level of 75%, the server determines the target value of this image as "minor crack" based on this confidence level.

[0137] Similarly, for images of bridge abutments, because the confidence level of the "moderate corrosion" category is 80%, which is significantly higher than other categories, the server identifies "moderate corrosion" as the target value for this image.

[0138] Suppose the server processes another image of the concrete surface of a dam slope. The main defect detection model gives a confidence level of 45% for "severe cracks," 40% for "large-area spalling," and 15% for "severe erosion." In this case, although the confidence level of "severe cracks" is not absolutely superior, it is relatively higher than the other two categories, and the server still identifies "severe cracks" as the target value to be determined.

[0139] The server performs this processing on a large number of images in the first sample set of concrete appearance images of buildings. By comprehensively analyzing the confidence scores given by the main defect detection model, a most likely target value is determined for each image from the target defect category set, providing an accurate reference for subsequent optimization and training of the secondary defect detection model.

[0140] In this embodiment of the invention, the defect detection model includes a first index vectorization unit and a first related element vectorization unit;

[0141] The present invention also provides the following embodiments.

[0142] The first sample building concrete appearance image is vectorized by the first index vectorization unit to obtain a first vector representation;

[0143] The first relevant element vectorization unit obtains a second vector representation of the first sample building concrete appearance image and the data in the first preset database;

[0144] Based on the deviation between the first vector representation and the second vector representation, knowledge transfer is performed on the defect detection model to obtain the apparent defect detection model.

[0145] In an embodiment of the present invention, for example, the server receives an image from a set of first sample building concrete appearance images. This is an image of the concrete surface of a large underground parking garage support, in which there are some minor cracks and local discoloration.

[0146] The server initiates processing of the image using the first indexed vectorization unit in the defect detection model. The first indexed vectorization unit converts various features in the image, such as the length, width, and direction of cracks, and the size and color value of discolored areas, into digital vectors.

[0147] Suppose that after processing, this parking lot pillar image is represented as a first vector with specific dimensions and values, such as a vector containing 512 values, each value representing a quantized feature of a certain aspect of the image.

[0148] The server continues to process other images in the first sample building concrete appearance image set in the same way, resulting in a series of first vector representations.

[0149] For the same image of an underground parking garage pillar, the server simultaneously initiates the vectorization unit of the first relevant element from the defect detection model.

[0150] The first relevant element vectorization unit not only considers the current image but also retrieves relevant data from a first preset database. This database may contain a large amount of previously collected feature descriptions, statistical information, and typical cases regarding concrete defects in buildings.

[0151] The first related element vectorization unit performs comprehensive analysis and vectorization processing on the features of the current image and related data in the database to obtain a second vector representation.

[0152] Suppose the second vector is a 1024-dimensional vector that incorporates the features of the current image and quantified information from relevant knowledge in the database.

[0153] The server performs the same operation on other images in the first sample building concrete appearance image set to obtain the corresponding second vector representation.

[0154] Taking this image of an underground parking garage pillar as an example, the server obtained its first vector representation and second vector representation.

[0155] The server begins calculating the deviation between the two vectors. The calculation of the deviation may involve various metrics such as the difference between vector elements, Euclidean distance, and cosine similarity.

[0156] Suppose the calculation results show a large deviation between the two vectors, which indicates that when the defect detection model processes this image, the output of its first index vectorization unit differs significantly from the output of the first related element vectorization unit, which incorporates database knowledge.

[0157] The server adjusts the parameters of the defect detection model based on this deviation. For example, if the first index vectorization unit does not quantize the crack features accurately enough, resulting in a large deviation from the second vector, the server will adjust the parameters related to crack feature extraction in that unit to enable more accurate image vectorization.

[0158] The server performs this deviation calculation and parameter adjustment for each image in the first sample set of building concrete appearance images.

[0159] Through multiple iterations and optimizations, the defect detection model gradually learns and improves, continuously reducing the deviation between the first and second vector representations. Ultimately, knowledge transfer is completed from the defect detection model, resulting in an appearance defect detection model capable of accurately detecting defects.

[0160] For example, for an image of a concrete tank wall in a wastewater treatment plant, an optimized appearance defect detection model can provide more accurate and reliable defect judgment results.

[0161] In this embodiment of the invention, the main defect detection model is trained using a multi-task learning method with the first sample building concrete appearance image set and a reinforcement learning method with the second sample building concrete appearance image set, which can be implemented through the following example.

[0162] Multi-task learning is performed on the feature extraction component based on the concrete appearance images of the first sample building in the first sample building concrete appearance image set.

[0163] By combining the gated recurrent component with the feature extraction component that has undergone multi-task learning, a defect detection model is obtained. The gated recurrent component is used to perform appearance defect detection in the target defect category set.

[0164] The defect detection model is obtained by performing reinforcement learning on the defect detection model using the second sample building concrete appearance image set and the sample target value.

[0165] In an embodiment of the invention, for example, the server acquires a series of images from a set of first sample building concrete appearance images. One of these images is of a concrete wall in the basement of a high-rise residential building, showing minor cracks and localized water stains.

[0166] The server begins using this image to perform multi-task learning on the feature extraction component. The tasks include identifying the type of crack (e.g., longitudinal cracks, transverse cracks), assessing the width of the cracks, and determining the area and depth of water stains.

[0167] The feature extraction component begins analyzing the image, extracting features relevant to each task. For example, to identify crack types, it focuses on the crack's direction and shape; to assess crack width, it extracts pixel information from the crack edges; and to determine the area and depth of water stains, it analyzes color variations and the continuity of the regions.

[0168] The server continues to use other first-sample images of the concrete surface of buildings, such as an image of the concrete surface of a bridge pier with signs of spalling and rust. While processing this image, the feature extraction component simultaneously learns to identify the degree of spalling, distinguish the type of rust, and performs multiple tasks.

[0169] By performing such multi-task learning on a large number of different first-sample building concrete appearance images, the feature extraction component gradually learns the ability to extract a variety of useful features from complex images.

[0170] After completing multi-task learning of the feature extraction component, the server combines the gated recurrent component with the trained feature extraction component.

[0171] The gated loop component is a component specifically designed for processing sequential data and performing dynamic predictions. In this scenario, it is used to detect and classify apparent defects within a target defect category set based on features extracted by the feature extraction component.

[0172] For example, for the image of the concrete wall of the basement of the high-rise residential building mentioned above, after the feature extraction component extracts the features of cracks and water stains, the gated loop component will determine the specific defect category of the image based on the sequence and combination of these features, such as the category of "minor cracks and moderate water stains".

[0173] Similarly, for images of bridge piers, the gated loop component combines the peeling and rust features extracted by the feature extraction component to provide corresponding defect detection results.

[0174] In this way, by combining the two components, the server obtains a preliminary defect detection model.

[0175] The server retrieved an image from the second set of concrete appearance images of the building. This was a severely damaged image of a large sluice gate concrete structure, showing obvious large cracks and large areas of concrete spalling. The sample target value clearly marked the specific type and severity of the defect.

[0176] The server inputs this image into the previously obtained defect detection model. After the model outputs its detection result, the server compares it with the sample target value. If the model's detection result is inconsistent with the sample target value, the server will penalize the model based on the deviation; if the detection result is close to or matches the sample target value, the server will reward the model.

[0177] For example, the model might initially classify the defects in this sluice gate image as "moderate cracks and moderate spalling," while the target value for the sample is "severe cracks and severe spalling." The server adjusts the model's parameters based on this discrepancy, enabling it to produce more accurate results the next time it encounters a similar image.

[0178] The server continues to use other images from the second set of images of the concrete appearance of buildings, such as an image of severely corroded concrete foundations of a chemical plant storage tank, to repeatedly reinforce the defect detection model and adjust its parameters.

[0179] Through reinforcement learning on a large number of second-sample building concrete appearance images with clear sample target values, the defect detection model is continuously optimized and improved. Finally, the server obtains a main defect detection model that can accurately detect various complex defects.

[0180] In this embodiment of the invention, the feature extraction component includes a second index vectorization unit and a second related element vectorization unit;

[0181] The multi-task learning of the feature extraction component based on the first sample building concrete appearance image in the first sample building concrete appearance image set can be implemented through the following example.

[0182] The first sample building concrete appearance image is vectorized by the second index vectorization unit to obtain a third vector representation;

[0183] The second related element vectorization unit obtains a fourth vector representation of the first sample building concrete appearance image and the data in the second preset database;

[0184] The feature extraction component performs multi-task learning based on the deviation between the third vector representation and the fourth vector representation.

[0185] In an embodiment of the present invention, for example, the server receives an image from a set of first sample building concrete appearance images. This is an image of the concrete floor of an old warehouse, in which there are multiple areas of cracks and wear.

[0186] The server initiates the second index vectorization unit in the feature extraction component to process the image. This unit analyzes and quantifies various features in the image. For example, it focuses on features such as the length, width, and distribution of cracks, and the area, shape, and roughness of worn regions.

[0187] After complex calculations and transformations, the second index vectorization unit converts these features into a digital vector, namely the third vector representation. This third vector is assumed to be a vector with 256 elements, each representing quantized information about a specific aspect of the image.

[0188] The server continues processing other images from the first set of building concrete appearance images, such as an image of concrete at a bridge expansion joint, showing obvious damage and deformation. Similarly, this image is transformed into a corresponding third vector representation using the second index vectorization unit.

[0189] For the image of the concrete floor of the old warehouse, the server simultaneously activates the second related element vectorization unit in the feature extraction component.

[0190] This unit not only analyzes the current image but also retrieves relevant data from a second pre-set database. This database may contain standard feature descriptions of various concrete defects, historical inspection data, material properties, and other information.

[0191] The second related element vectorization unit integrates the features of the current image with the data in the database and transforms it into a digital vector, namely the fourth vector representation. This fourth vector is assumed to be a 512-dimensional vector, which incorporates the features of the current image and quantified information from relevant knowledge in the database.

[0192] The server performs the same operation on other images in the first sample building concrete appearance image set to obtain the corresponding fourth vector representation.

[0193] Taking the image of the concrete floor of an old warehouse as an example, the server obtained its third and fourth vector representations.

[0194] The server begins calculating the deviation between the two vectors. The degree of deviation can be quantified, for example, by calculating the difference between vector elements, Euclidean distance, or cosine similarity.

[0195] Assuming the calculation results show a large deviation, this means that there is a significant difference between the feature representation obtained by the second index vectorization unit processing the image alone and the feature representation obtained by the second related element vectorization unit processing the image in combination with database information.

[0196] The server adjusts the parameters of the feature extraction component based on this bias to optimize its performance in multi-task learning. For example, if the bias indicates that the second index vectorization unit is not accurately quantizing the crack length, the server will adjust the relevant parameters to enable it to extract crack length features more precisely.

[0197] The server also performs similar deviation calculations and parameter adjustments for the concrete images at the bridge expansion joints.

[0198] By performing such bias calculations and parameter optimizations on a large number of images in the first sample set of concrete appearance images of buildings, the feature extraction component can continuously learn and improve, better adapting to the requirements of multi-task learning.

[0199] For example, for a concrete image of a water tower foundation, the optimized feature extraction component can more accurately extract various defect features such as cracks, corrosion, and deformation, providing a more reliable basis for subsequent defect detection and classification.

[0200] In this embodiment of the invention, the combination of the gated loop component with the feature extraction component that has undergone multi-task learning to obtain the defect detection model can be implemented through the following example.

[0201] The gated loop component is concatenated with the second index vectorization unit, which has undergone multi-task learning, to obtain the defect detection model.

[0202] In an embodiment of the present invention, for example, in a large building quality monitoring center, the server is undertaking the important task of building a defect detection model.

[0203] The server has completed multi-task learning of the feature extraction component, in which the second index vectorization unit has the ability to extract key features from the concrete appearance image of the first sample building and quantize them.

[0204] Taking an image of the concrete surface of a hydroelectric dam as an example, the second index vectorization unit, after multi-task learning, can transform the defect features of the dam surface, such as cracks, holes, and spalling, into accurate digital vector representations.

[0205] Next, the server will cascade the gated loop component with the second index vectorization unit, which has been learned through multi-task training, to build a defect detection model.

[0206] The gated loop component acts like an intelligent decision-making center. It can receive the vectors output by the second index vectorization unit and perform in-depth analysis and reasoning based on the information in these vectors.

[0207] The server first tightly connects the output of the second indexed vectorization unit to the input of the gated loop component. This way, when a new concrete appearance image is input, the vector generated by the second indexed vectorization unit after processing it can be seamlessly passed to the gated loop component.

[0208] For example, for an image of a concrete pillar at a port dock, the second index vectorization unit converts features such as fine cracks and erosion marks on the pillar surface into vectors, which are then immediately fed into the gating loop component.

[0209] Once the gated loop component receives these vectors, it begins to perform its unique function. It memorizes and updates the input vector sequence, capturing long-term dependencies and dynamic changes.

[0210] Suppose that in this image of a concrete pillar at a port dock, the characteristics of cracks develop in different locations and at different times in a certain pattern. The gated loop component can analyze the input vector sequence to discover this pattern and make a more accurate judgment on the type, severity and development trend of the defect.

[0211] The server continues to process other types of concrete appearance images of buildings, such as an image of the concrete piers of an overpass. The second index vectorization unit, after multi-task learning, extracts the defect feature vectors of the pier surface and then passes them to the gated recurrent component.

[0212] By comprehensively analyzing these continuously input vectors, the gated loop component can determine whether the defects on the pier surface are due to crack expansion caused by long-term vehicle vibration or concrete spalling caused by environmental erosion.

[0213] Through this cascading approach, the server successfully combined the second index vectorization unit and the gated loop component to form a model that can efficiently and accurately detect apparent defects in building concrete.

[0214] For example, given an image of a concrete column in the underground parking lot of a large shopping mall, this defect detection model can quickly provide a detailed report on the surface defects of the column, including the specific type, location, size of the defect, and its potential impact on the structural safety of the building.

[0215] After continuous testing and optimization using concrete surface images of buildings under various scenarios and conditions, the server finally obtained a high-performance and highly reliable defect detection model, providing strong technical support for ensuring the safety and stability of buildings.

[0216] This invention provides a computer device 100, which includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the aforementioned artificial intelligence-based method for detecting apparent defects in underwater structure concrete. Figure 2 As shown, Figure 2 This is a structural block diagram of a computer device 100 provided in an embodiment of the present invention. The computer device 100 includes a memory 111, a processor 112, and a communication unit 113.

[0217] To enable data transmission or interaction, the memory 111, processor 112, and communication unit 113 are electrically connected to each other, either directly or indirectly. For example, these components can be electrically connected to each other via one or more communication buses or signal lines.

[0218] For illustrative purposes, the foregoing description has been made with reference to specific embodiments. However, the foregoing illustrative discussions are not intended to be exhaustive or to limit the present disclosure to the precise forms disclosed. Numerous modifications and variations are possible in accordance with the foregoing teachings. These embodiments were chosen and described in order to best illustrate the principles of the present disclosure and its practical application, thereby enabling those skilled in the art to best utilize the disclosure and to employ various embodiments with different modifications to suit a particular intended application.

Claims

1. An underwater building concrete surface defect detection method based on artificial intelligence, characterized in that, The method comprises the following steps: Air calibration and underwater calibration are performed on a preset color camera and a preset black-and-white camera; A first image and a second image of a concrete surface of an underwater building are obtained by the calibrated color camera, the first image having laser stripes and the second image not having laser stripes, the laser stripes being generated based on underwater refraction; A third image of the concrete surface of the underwater building is obtained by the calibrated black-and-white camera; The first image and the third image are matched based on the laser stripes, and on the basis of successful matching, a point cloud image of the concrete surface of the underwater building is constructed based on the first image and the third image, and color enhancement is performed based on the second image to obtain a target point cloud image of the concrete surface of the underwater building; Defect recognition is performed on the target point cloud image of the concrete surface of the underwater building based on an apparent defect detection model to obtain a defect detection result corresponding to the target point cloud image of the concrete surface of the underwater building; The defect recognition performed on the target point cloud image of the concrete surface of the underwater building based on the apparent defect detection model to obtain the defect detection result corresponding to the target point cloud image of the concrete surface of the underwater building comprises the following steps: A first sample building concrete surface image set and a second sample building concrete surface image set are obtained, the first sample building concrete surface image set comprising a first sample building concrete surface image, the second sample building concrete surface image set comprising a second sample building concrete surface image set configured with sample target values, and the second sample building concrete surface image set being classified into a target defect category set; A master defect detection model is trained in a multi-task learning method through the first sample building concrete surface image and in a reinforcement learning manner through the second sample building concrete surface image set; A slave defect detection model is obtained, the slave defect detection model being an original model whose model parameters have not been trained; Knowledge of the model parameters of the slave defect detection model is transferred to the slave defect detection model through the first sample building concrete surface image and the master defect detection model as a teacher model, so as to obtain an apparent defect detection model; Defect recognition is performed on a target point cloud image of a concrete surface of an underwater building through the apparent defect detection model to obtain a defect detection result of the target point cloud image of the concrete surface of the underwater building in the target defect category set; The training of the master defect detection model in a multi-task learning method through the first sample building concrete surface image and in a reinforcement learning manner through the second sample building concrete surface image set comprises the following steps: Multi-task learning is performed on a feature extraction component based on the first sample building concrete surface image in the first sample building concrete surface image set; A defect detection model is obtained by combining a gated recurrent component with the feature extraction component that has been subjected to multi-task learning, the gated recurrent component being used for apparent defect detection in the target defect category set. The defect detection model is reinforced by the second sample building concrete apparent image set in the second sample building concrete apparent image set and the sample target value, and a main defect detection model is obtained. The feature extraction component includes a second index vectorization unit and a second related element vectorization unit. The multi-task learning of the feature extraction component based on the first sample building concrete apparent image in the first sample building concrete apparent image set includes: The first sample building concrete apparent image is vectorized by the second index vectorization unit to obtain a third vector representation; The fourth vector representation of the first sample building concrete apparent image and the data in the second preset database is obtained by the second related element vectorization unit; The multi-task learning of the feature extraction component based on the deviation between the third vector representation and the fourth vector representation.

2. The method of claim 1, wherein, The model parameters of the slave defect detection model are knowledge transferred by taking the main defect detection model as a teacher model based on the first sample building concrete apparent image, and an apparent defect detection model is obtained. The first sample building concrete apparent image in the first sample building concrete apparent image set is defect-recognized by the main defect detection model to obtain a pending target value corresponding to the class in the target defect class set of the first sample building concrete apparent image. The first sample building concrete apparent image in the first sample building concrete apparent image set is defect-recognized by the slave defect detection model to obtain a recognition result corresponding to the first sample building concrete apparent image. The model parameters of the slave defect detection model are optimized based on the deviation between the pending target value and the recognition result, and the apparent defect detection model is obtained.

3. The method of claim 2, wherein, The model parameters of the slave defect detection model are optimized based on the deviation between the pending target value and the recognition result, and the apparent defect detection model is obtained. The credibility index of the pending target value is obtained. The deviation between the pending target value and the recognition result under the credibility index is determined, and the model parameters of the slave defect detection model are optimized based on the deviation, and the apparent defect detection model is obtained.

4. The method of claim 2, wherein, The first sample building concrete apparent image in the first sample building concrete apparent image set is defect-recognized by the main defect detection model to obtain a pending target value corresponding to the class in the target defect class set of the first sample building concrete apparent image. The first sample building concrete apparent image in the first sample building concrete apparent image set is defect-recognized by the main defect detection model to obtain a confidence of the first sample building concrete apparent image corresponding to the classification in the target defect class set. The pending target value corresponding to the first sample building concrete apparent image is determined from the target defect class set based on the confidence.

5. The method of claim 2, wherein, The first index vectorization unit and the first related element vectorization unit are included in the defect detection model; The method further comprises: vectorizing the first sample building concrete apparent image through the first index vectorization unit to obtain a first vector representation; obtaining a second vector representation of vectorizing the first sample building concrete apparent image and data in a first preset database by the first related element vectorization unit; performing knowledge transfer on the defect detection model based on the deviation between the first vector representation and the second vector representation to obtain the apparent defect detection model.

6. The method of claim 1, wherein, The defect detection model is obtained by combining the gated recurrent component with the feature extraction component subjected to multi-task learning, comprising: The gated recurrent component is cascaded with the second index vectorization unit subjected to multi-task learning to obtain the defect detection model.

7. A server system, characterized by The server is used to execute the method in any one of claims 1-6.

Images