A visual algorithm for underwater pipeline damage inspection based on 3D reconstruction and deep learning

Abstract

The present invention discloses a visual algorithm for underwater pipeline damage inspection based on 3D reconstruction and deep learning. The algorithm comprises the following steps: Step 1: Using an underwater camera to capture and preprocess video or image data of the pipeline underwater; Step 2: Calibrate the underwater camera and estimate the intrinsic and extrinsic parameters of the camera by minimizing the reprojection error; Step 3: Generate a 3D model of the underwater pipeline; Step 4: Segment the pipeline from the surrounding environment in the 3D model, extract the pipeline area, and annotate the extracted pipeline area and the point cloud dataset of the defective area to obtain pipeline annotation data, and classify the defects; Step 5: Use a pre-trained convolutional neural network deep learning model to perform damage inspection and analysis on the segmented pipeline; Step 6: Visually display the analysis results. This method addresses the problems of low efficiency, high cost, and high safety risks of traditional underwater pipeline inspection methods.

Description

Technical Field

[0001] The present invention relates to the field of image processing and computer vision technology, and specifically to an underwater pipeline damage inspection visual algorithm based on three-dimensional reconstruction and deep learning. Background Art

[0002] Traditional methods for underwater pipeline inspection and maintenance suffer from several flaws and technical issues. First, traditional underwater pipeline inspection typically relies on manual patrols or visual inspections by divers, a method that is inefficient, costly, and poses safety risks. Second, traditional image processing technology has limited ability to handle issues such as poor image quality and uneven lighting in underwater environments, making it difficult to accurately detect and analyze pipeline damage. Furthermore, traditional pipeline damage detection methods often only provide simple information on damage location and size, lacking a more detailed analysis of damage type and extent, hindering the accuracy and efficiency of pipeline maintenance and repair.

[0003] To address these technical challenges, we have proposed a visual algorithm for underwater pipeline damage inspection based on 3D reconstruction and deep learning. By leveraging advanced 3D reconstruction technology and deep learning algorithms, combined with underwater camera equipment and image processing techniques, we achieve efficient and accurate inspection and analysis of underwater pipelines. This algorithm automates the detection and analysis of pipeline damage, improving accuracy and efficiency while reducing labor costs and safety risks. Furthermore, the algorithm provides a detailed analysis of pipeline damage, including information such as damage location, type, and size, providing a more accurate reference for pipeline maintenance and repair. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention proposes a visual algorithm for underwater pipeline damage inspection based on three-dimensional reconstruction and deep learning to solve the problems of low efficiency, high cost and high safety risks of traditional underwater pipeline inspection methods.

[0005] In order to solve the above technical problems, the technical solution of the present invention is:

[0006] A visual algorithm for underwater pipeline damage inspection based on 3D reconstruction and deep learning includes the following steps:

[0007] Step 1: Use underwater camera equipment to shoot video or image data of the pipeline underwater and pre-process it;

[0008] Step 2: Calibrate the underwater camera and estimate the intrinsic and extrinsic parameters of the camera by minimizing the reprojection error.

[0009] Step 3: Generate a three-dimensional model of the underwater pipeline;

[0010] Step 3.1: Use the SURF algorithm to extract features from the preprocessed image data. The extracted features include key points and feature descriptors in the image.

[0011] Step 3.2: Using the key points and feature descriptor data of each image obtained in step 3.1, the three-dimensional coordinates of the feature points in the image are calculated using triangulation methods to generate sparse three-dimensional point cloud data, thereby obtaining a point cloud dataset. The three-dimensional point cloud data is then converted to the world coordinate system based on the intrinsic and extrinsic parameters.

[0012] Step 3.3: Based on the sparse 3D point cloud data, a dense 3D point cloud is generated using a cubic spline interpolation algorithm to obtain a 3D model.

[0013] Step 4: Segment the pipeline and the surrounding environment in the 3D model, extract the pipeline area, annotate the extracted pipeline area and the point cloud dataset of the defective area to obtain pipeline annotation data, and classify the defects;

[0014] Step 5: Use the pre-trained convolutional neural network deep learning model to perform damage analysis on the segmented pipeline;

[0015] Step 6: Display the analysis results in a visual way.

[0016] Preferably, in step 1, the preprocessing method includes removing noise in the underwater environment, adjusting image quality, image denoising, image enhancement and image registration.

[0017] Preferably, the image registration method is:

[0018] First, determine the key points by finding the extreme points in the DoG pyramid. The method is as follows:

[0019] D(x,y,σ)=D(x,y,σ)-max(D(x,y-1,σ),D(x,y+1,σ),D(x-1,y,σ),D(x+1,y,σ),D(x,y,σ-1),D(x,y,σ+1))

[0020] Where σ is the scale parameter, x, y are the image coordinates;

[0021] Then determine the main direction of the key point so that the key point has rotation invariance, as follows:

[0022]

[0023] Where D(x,y) is the DoG response and θ(x,y) is the direction of the keypoint.

[0024] Preferably, the calibration method in step 2 is:

[0025] Step 2.1, obtain a calibration plate with regular grid points;

[0026] Step 2.2: Fix the calibration plate in a known position so that the calibration plate is completely visible in the field of view of the camera.

[0027] Step 2.3, adjust the position and attitude parameters of the calibration plate in the world coordinate system through the rotation matrix R and translation vector t, where R is a 3x3 rotation matrix and t is a 3x1 translation vector;

[0028] Step 2.4: Collect images of the calibration plate at multiple angles and positions. Assume there are N images, each of which corresponds to a point in the world coordinate system and a point in the pixel coordinate system, which can be expressed as:

[0029] in is a point in the world coordinate system, is a point in pixel coordinates.

[0030] Preferably, in step 3, the method of extracting features using the SURE algorithm is:

[0031] First, obtain the Hessian matrix using:

[0032]

[0033] Then use the Hessian matrix to detect the points of interest in the image, where L xx (x,σ),L xy (x,σ),L yy (x,σ) are the second-order derivatives of the image at point x and scale σ respectively;

[0034] After the points of interest are determined, the integral image of the Haar wavelet response is used to determine the principal directions, which are determined by the following formula:

[0035]

[0036] Where HaarX(p) and HaarY(p) are the Haar wavelet responses at p;

[0037] Finally, a gradient-based descriptor is used to describe the image area around the interest point to obtain the feature descriptor.

[0038] Preferably, the descriptor is calculated by dividing the area around the point of interest into small sub-areas, calculating the Haar wavelet response for each sub-area, and constructing a vector, connecting the vectors of all sub-areas to form the final descriptor.

[0039] Preferably, in step 3, the method for generating sparse three-dimensional point cloud data is:

[0040] P=s·(A -1 a1-B -1 b1)

[0041] Where P represents the three-dimensional coordinates of the feature point, s is the scale factor, A and B are the camera intrinsic parameter matrices, and a1 and b1 are the normalized plane coordinates of the feature point under two viewing angles;

[0042] The method to generate dense 3D point cloud data is:

[0043] f(x)=a0+a1(xx i )+a2(xx i ) 2 +a3(xx i ) 3

[0044] Among them, a0, a1, a2, and a3 are coefficients calculated from known points and derivatives.

[0045] Preferably, in step 4, when segmenting the three-dimensional model to extract the pipeline area, the edge of the pipeline is judged using curvature, and the formula is as follows:

[0046]

[0047] Where n is the normal vector, p represents the three-dimensional coordinates of the feature point, Represents the gradient.

[0048] Preferably, in step 5, the point cloud dataset obtained in step 3 and the pipeline annotation data obtained in step 4 are used as input to train a convolutional neural network deep learning model.

[0049] Preferably, the convolutional neural network deep learning model training process is expressed as a process of optimizing a loss function, and the loss function expression is as follows:

[0050]

[0051] Where N is the number of training samples, C is the number of categories, and y i,c is the actual label of the i-th sample (0 or 1, to indicate whether it is damaged), is the model's predicted probability of category c for the i-th sample.

[0052] The present invention has the following characteristics and beneficial effects:

[0053] This technical solution, utilizing advanced 3D reconstruction technology and deep learning algorithms, combined with underwater camera equipment and image processing techniques, avoids the errors often associated with conventional monocular image recognition algorithms due to interference from low light levels and impurities. Furthermore, the algorithm enables detailed analysis of pipeline damage, including 3D information such as damage location, type, and size. This provides a more comprehensive reference for pipeline maintenance and repair, reducing both the cost and difficulty of repair. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0055] Figure 1 This is a flowchart of the three-dimensional reconstruction of the present invention.

[0056] Figure 2 This is a flow chart of pipeline segmentation and damage analysis of the present invention.

[0057] Figure 3 Flowchart for visual display of results of the present invention.

[0058] Figure 4 A comparison chart of the present invention and conventional algorithms. DETAILED DESCRIPTION

[0059] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0060] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, features defined as "first", "second", etc. may explicitly or implicitly include one or more of the features. In the description of the present invention, unless otherwise specified, "multiple" means two or more.

[0061] In the description of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "mounted," "connected," and "connected" should be understood in a broad sense. For example, they may refer to fixed connections, detachable connections, or integral connections; mechanical connections or electrical connections; direct connections or indirect connections through an intermediate medium; and internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0062] The present invention provides a visual algorithm for underwater pipeline damage inspection based on 3D reconstruction and deep learning. Figure 1 and Figure 2 As shown, the following steps are included

[0063] Step 1: Use underwater camera equipment to capture video or image data of the pipeline underwater.

[0064] Specifically, in this embodiment, the following steps are included:

[0065] Step 1.1: Turn on the lights of the underwater camera (using an underwater robot as an example below) and enable line patrol mode, and move along the pipeline;

[0066] Step 1.2: Collect pipeline-related video information along the way and store it in an SD card;

[0067] Step 1.3: After collecting enough information, open the underwater camera, remove the SD card, read it, and transfer it to the computer for subsequent processing.

[0068] Step 2: Preprocess the collected data, including removing noise in the underwater environment, adjusting image quality, image denoising, image enhancement, and image registration.

[0069] Specifically, in this embodiment, image denoising uses appropriate image processing techniques, such as median filtering and Gaussian filtering, to remove noise from the image and improve image quality. Image enhancement performs image enhancement processing, such as adjusting contrast and brightness, to improve image clarity and quality.

[0070] A further configuration of this embodiment is image registration, which performs registration on the multiple images collected to ensure that images at different angles or positions can be aligned.

[0071] For image registration, the following steps are included

[0072] Find extreme points in the DoG pyramid to determine key points

[0073] D(x,y,σ)=D(x,y,σ)-max(D(x,y-1,σ),D(x,y+1,σ),D(x-1,y,σ),D(x+1,y,σ),D(x,y,σ-1),D(x,y,σ+1))

[0074] Where σ is the scale parameter and x,y are the image coordinates.

[0075] Determine the main direction of the key point so that the key point has rotation invariance.

[0076]

[0077] Where D(x,y) is the DoG response and θ(x,y) is the direction of the keypoint.

[0078] In this embodiment, the key point is described as a 128-dimensional vector, including gradient histograms in 8 directions of 16 sub-regions around the key point.

[0079] Step 3: Calibrate the underwater camera equipment to determine the camera's intrinsic and extrinsic parameters.

[0080] The following steps are involved:

[0081] Step 3.1: Select a calibration plate suitable for underwater camera equipment, usually a calibration plate with a regular grid, to calibrate the camera parameters.

[0082] Step 3.2: Camera installation and calibration

[0083] Step 3.2.1: Fix the calibration plate in a known position to ensure that the calibration plate appears completely in the camera's field of view. It is necessary to consider the position and attitude parameters of the calibration plate in the world coordinate system, which are usually represented by the rotation matrix R and the translation vector t, where R is a 3x3 rotation matrix and t is a 3x1 translation vector

[0084]

[0085] Step 3.2.2: Collect images of the calibration plate at multiple angles and positions: Assume there are N images, each image corresponds to a point in the world coordinate system and a point in the pixel coordinate system, which can be expressed as:

[0086] in is a point in the world coordinate system, is a point in pixel coordinates

[0087] Step 3.2.3: Use the calibration algorithm to calculate the camera's intrinsic and extrinsic parameters: Zhang's calibration algorithm estimates the camera's intrinsic and extrinsic parameters by minimizing the reprojection error. Its core formula is:

[0088] Where s is the scale factor, f x and f y is the focal length, c x and c y is the principal point coordinate, r ij is the element of the rotation matrix, t i are the elements of the translation vector

[0089] In order to further illustrate the effectiveness of this embodiment, this embodiment also proposes to evaluate the calibration results:

[0090] The calibration results are evaluated using the reprojection error as the evaluation criterion to check the accuracy and stability of the intrinsic and extrinsic parameters. The formula is as follows: Where N is the number of calibration points on the calibration plate, (u_i', vi_i') is the projection position of the calibration point on the image calculated by internal and external parameters, and (u_i, vi_i) is the calibration point position in the actual image.

[0091] Step 4: Use the calibrated camera parameters to perform 3D reconstruction on the preprocessed image data. Generate a 3D model of the underwater pipeline, including information such as the shape and size of the pipeline.

[0092] Specifically, the following steps are included:

[0093] Step 4.1: Use the SURF algorithm to perform feature extraction on the preprocessed image data to extract key points and feature descriptors in the image.

[0094] First, obtain the Hessian matrix using:

[0095]

[0096] Then use the Hessian matrix to detect the points of interest in the image, where L xx (x,σ),L xy (x,σ),L yy (x,σ) are the second-order derivatives of the image at point x and scale σ, respectively.

[0097] After the points of interest are identified, the SURF algorithm uses the integral image of the Haar wavelet response to determine the principal directions.

[0098] The main directions are determined by the following formula:

[0099]

[0100] Where HaarX(p) and HaarY(p) are the Haar wavelet responses at p. The SURF algorithm uses a gradient-based descriptor to describe the image region around the point of interest.

[0101] Furthermore, descriptor calculation includes the following steps: The area around the point of interest is divided into small subregions. In this embodiment, 4x4 subregions are used. A Haar wavelet response is calculated for each subregion, and a vector is constructed. The vectors of all subregions are concatenated to form the final descriptor.

[0102] Step 4.2: Based on the feature matching results, use triangulation to calculate the 3D coordinates of the feature points in the image and generate sparse 3D point cloud data. Based on the camera's intrinsic and extrinsic parameters, convert the 3D point cloud data into the world coordinate system.

[0103] Step 4.2.1: Triangulate to obtain sparse point cloud. The formula is P = s (A -1 a1-B -1 b1), where P represents the three-dimensional coordinates of the feature point, s is the scale factor, A and B are the camera intrinsic parameter matrices, and a1 and b1 are the normalized plane coordinates of the feature point under two viewing angles.

[0104] Step 4.2.2: Use the camera's internal and external parameters to perform point cloud correction. The correction formula is P world =R·P camera +t, where P world Represents the three-dimensional point cloud coordinates in the world coordinate system, R and t represent the camera's rotation matrix and translation vector, P camera Represents the 3D point cloud coordinates in the camera coordinate system.

[0105] Step 4.3: Based on the sparse 3D point cloud data, use the cubic spline interpolation algorithm to generate a dense 3D point cloud to improve the accuracy and completeness of the reconstruction. The formula is:

[0106] f(x)=a0+a1(xx i )+a2(xx i ) 2 +a3(xx i ) 3 ,

[0107] Where a0, a1, a2, and a3 are coefficients calculated from known points and derivatives.

[0108] Step 5: Segment the pipeline from its surroundings in the 3D model to extract the pipeline area. Label the point cloud datasets for the entire pipeline and the defective portion, and classify the defects for subsequent model training.

[0109] It should be noted that when segmenting, the edge of the pipe is judged using curvature, and the formula is as follows: Where n is the normal vector, p is the point cloud coordinate, Represents the gradient.

[0110] Step 6: Use a convolutional neural network (CNN) deep learning model to perform damage analysis on the segmented pipeline.

[0111] Specifically, the following steps are included

[0112] Step 6.1: Train a convolutional neural network (CNN) deep learning model based on the point cloud dataset obtained in step 4 and the pipeline annotation data obtained in step 5. The CNN training process can be expressed as a process of optimizing a loss function. Where N is the number of training samples, C is the number of categories, and y i,c is the actual label of the i-th sample (0 or 1, to indicate whether it is damaged), is the model's predicted probability of category c for the i-th sample.

[0113] Step 6.3: Perform predictions on new pipeline images and use the trained deep learning model to detect pipeline damage and cracks. The prediction results can provide information such as damage location, type, and size.

[0114] 6.4: Based on the model prediction results, perform damage analysis to determine the damage to the pipeline, including information such as the location, type, and size of the damage. The damage analysis results can provide a reference for pipeline maintenance and repair, thereby improving the service life and safety of the pipeline.

[0115] Furthermore, to ensure the effectiveness of the convolutional neural network (CNN) deep learning model, this embodiment also provides an evaluation of the model prediction results, calculates the model's accuracy, recall rate and other indicators, and evaluates the model's performance. The model can be evaluated through methods such as cross-validation to improve the model's generalization ability. The threshold uses the F1 score (F1 Score), which is an indicator that comprehensively considers precision and recall. It is the harmonic mean of precision and recall:

[0116] Step 7: Display the analysis results in a visual way, including information such as damage location, type, size, etc. Provide intuitive result display to help users understand the damage status of the pipeline. Figure 3 and 4As shown, the result of this process is a three-dimensional rotatable and scalable model, in which white points represent key points of healthy and normal pipelines, black points represent damaged and corroded points, and finally a red box represents the damaged area and the severity of the damage. This is essentially different from the output of conventional visual processing. Conventional algorithms only perform image processing on two-dimensional pixels, and the output is a series of two-dimensional coordinates that are general and not very accurate. For the results of damage identification, it is usually still necessary to confirm the original image with the naked eye to judge the general situation. However, due to the use of binocular recognition and deep learning, the accuracy and reliability of this algorithm have been greatly improved. It can basically accurately provide the three-dimensional area of ​​the damaged area, and the effect is better than naked eye recognition. Therefore, the output results can serve as an important basis for pipeline maintenance personnel to plan maintenance projects. See the specific example effect diagram. Figure 4 .

[0117] Step 8: Store the analysis results for later reference and use.

[0118] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. It will be apparent to those skilled in the art that various changes, modifications, substitutions, and variations of these embodiments, including components, without departing from the principles and spirit of the present invention are still within the scope of protection of the present invention.

Claims

1. A visual algorithm for underwater pipeline damage inspection based on 3D reconstruction and deep learning, characterized by: The steps include: Step 1: Use underwater camera equipment to shoot video or image data of the pipeline underwater and pre-process it; Step 2: Calibrate the underwater camera and estimate the intrinsic and extrinsic parameters of the camera by minimizing the reprojection error. Step 3: Generate a three-dimensional model of the underwater pipeline; Step 3.1: Use the SURF algorithm to extract features from the preprocessed image data. The extracted features include key points and feature descriptors in the image. The method of extracting features using the SURF algorithm is: First, obtain the Hessian matrix using: Then use the Hessian matrix to detect the points of interest in the image, where L xx (x,σ),L xy (x,σ),L yy (x,σ) are the second-order derivatives of the image at point x and scale σ respectively; After the points of interest are determined, the integral image of the Haar wavelet response is used to determine the principal directions, which are determined by the following formula: Where HaarX(p) and HaarY(p) are the Haar wavelet responses at p; Finally, a gradient-based descriptor is used to describe the image area around the interest point to obtain a feature descriptor; The descriptor is calculated by dividing the area around the point of interest into small sub-regions, calculating the Haar wavelet response for each sub-region, and constructing a vector, connecting the vectors of all sub-regions to form the final descriptor; Step 3.2: Using the key points and feature descriptor data of each image obtained in step 3.1, the three-dimensional coordinates of the feature points in the image are calculated using triangulation methods to generate sparse three-dimensional point cloud data, thereby obtaining a point cloud dataset. The three-dimensional point cloud data is then converted to the world coordinate system based on the intrinsic and extrinsic parameters. Step 3.3: Based on the sparse 3D point cloud data, a dense 3D point cloud is generated using a cubic spline interpolation algorithm to obtain a 3D model. Step 4: Segment the pipeline and the surrounding environment in the 3D model, extract the pipeline area, annotate the extracted pipeline area and the point cloud dataset of the defective area to obtain pipeline annotation data, and classify the defects; When segmenting the 3D model to extract the pipeline area, the edge of the pipeline is determined using curvature. The formula is as follows: Where n is the normal vector, P is the three-dimensional coordinate of the feature point, represents the gradient; Step 5: Use the pre-trained convolutional neural network deep learning model to perform damage analysis on the segmented pipeline; Step 6: Display the analysis results in a visual way.

2. The underwater pipeline damage detection visual algorithm based on three-dimensional reconstruction and deep learning according to claim 1 is characterized in that: In step 1, the preprocessing method includes removing noise in the underwater environment, adjusting image quality, image denoising, image enhancement and image registration.

3. The underwater pipeline damage detection visual algorithm based on three-dimensional reconstruction and deep learning according to claim 2 is characterized in that: The image registration method is: First, determine the key points by finding the extreme points in the DoG pyramid. The method is as follows: D(x,y,σ)=D(x,y,σ)-max(D(x,y-1,σ),D(x,y+1,σ),D(x-1,y,σ),D(x+1,y,σ),D(x,y,σ-1),D(x,y,σ+1)) Where σ is the scale parameter, x, y are the image coordinates; Then determine the main direction of the key point so that the key point has rotation invariance, as follows: Where D(x,y) is the DoG response and θ(x,y) is the direction of the keypoint.

4. The underwater pipeline damage detection visual algorithm based on three-dimensional reconstruction and deep learning according to claim 1 is characterized in that: The calibration method in step 2 is: Step 2.1, obtain a calibration plate with regular grid points; Step 2.2: Fix the calibration plate in a known position so that the calibration plate is completely visible in the field of view of the camera. Step 2.3, adjust the position and attitude parameters of the calibration plate in the world coordinate system through the rotation matrix R and translation vector t, where R is a 3x3 rotation matrix and t is a 3x1 translation vector; Step 2.4: Collect images of the calibration plate at multiple angles and positions. Assume there are N images, each of which corresponds to a point in the world coordinate system and a point in the pixel coordinate system, which can be expressed as: in is a point in the world coordinate system, is a point in pixel coordinates.

5. The underwater pipeline damage detection visual algorithm based on three-dimensional reconstruction and deep learning according to claim 1 is characterized in that: In step 3, the method for generating sparse three-dimensional point cloud data is: P=s·(A -1 a1-B -1 b1) Where P represents the three-dimensional coordinates of the feature point, s is the scale factor, A and B are the camera intrinsic parameter matrices, and a1 and b1 are the normalized plane coordinates of the feature point under two viewing angles; The method to generate dense 3D point cloud data is: f(x)=a0+a1(x-x i )+a2(x-x i ) 2 +a3(x-x i ) 3 Among them, a0, a1, a2, and a3 are coefficients calculated from known points and derivatives.

6. The underwater pipeline damage detection visual algorithm based on three-dimensional reconstruction and deep learning according to claim 1 is characterized in that: In step 5, the point cloud dataset obtained in step 3 and the pipeline annotation data obtained in step 4 are used as input to train a convolutional neural network deep learning model.

7. The underwater pipeline damage detection visual algorithm based on three-dimensional reconstruction and deep learning according to claim 6 is characterized in that: The convolutional neural network deep learning model training process is expressed as a process of optimizing the loss function, and the loss function expression is as follows: Where N is the number of training samples, C is the number of categories, and y i,c is the actual label of the i-th sample, y i,c It is 0 or 1 to indicate whether it is damaged. is the model's predicted probability of category c for the i-th sample.

Images