Insulator surface defect detection method and system based on photometric stereo vision

Abstract

The present invention discloses a method and system for detecting insulator surface defects based on photometric stereo vision. The method comprises the following steps: S1. acquiring multiple insulator surface images obtained using photometric stereo; S2. restoring a 2.5D depth map of the insulator surface from the insulator surface images; S3. denoising the 2.5D depth map of the insulator surface; S4. performing key point extraction and surface fitting on the denoised depth map to obtain key control points and corresponding surfaces as sample set A; constructing random control point coordinate combinations and corresponding surfaces as sample set B; training a feature extraction network using sample sets A and B to obtain a control point extractor to obtain control points of the surface and obtain an insulator surface model; and S5. calculating parameter information of the insulator surface based on the insulator surface model, and performing defect detection based on the parameter information. The present invention has the advantages of high detection accuracy.

Description

Technical Field

[0001] The present invention mainly relates to the technical field of insulator detection, and in particular to a method and system for detecting surface defects of insulators based on photometric stereo vision. Background Art

[0002] Deformation of pot-shaped insulators during installation, excessive strain at connections, or damage and partial detachment due to aging can impact the safety and stability of transmission lines. Due to the appearance and material properties of pot-shaped insulators, inspection using TOF cameras or LiDAR can easily reveal problems such as voids and insufficient stability in close-range measurements. When inspecting threaded objects using conventional monocular vision systems, variations in thread height create noticeable shadows, making visual inspection of defects extremely difficult. Summary of the Invention

[0003] In view of the technical problems existing in the prior art, the present invention provides a method and system for detecting surface defects of insulators based on photometric stereo vision, which improve the detection accuracy.

[0004] In order to solve the above technical problems, the technical solution proposed by the present invention is:

[0005] A method for detecting surface defects of insulators based on photometric stereo vision comprises the following steps:

[0006] S1. Acquire multiple insulator surface images obtained by photometric stereo;

[0007] S2. Restore a 2.5D depth map of the insulator surface from the insulator surface image;

[0008] S3. Denoise the 2.5D depth map of the insulator surface to obtain a denoised depth map;

[0009] S4. Extract key points and perform surface fitting on the denoised depth map to obtain key control points and corresponding surfaces as sample set A; construct random control point coordinate combinations and corresponding surfaces as sample set B;

[0010] Through the sample set A and sample set B based on The feature extraction network is trained to obtain the final control point extractor to obtain the control points of the surface, and then the surface model of the insulator is obtained;

[0011] S5. Based on the insulator surface model, calculate parameter information of the insulator surface and perform defect detection based on the parameter information.

[0012] Preferably, in step S2, a mapping relationship is established between the pixel value of the single-sided fill light image and the unit surface normal vector of the 2.5D image of the insulator surface corresponding to the positive projection of the corresponding fill light source according to the Lambertian reflection model. Combined with the light source direction matrix and light source intensity of multiple groups of light sources, the normal vector set is solved by the least squares method, and normalized to generate a 2.5D depth map representing the surface depth, that is, a depth map matrix.

[0013] Preferably, the specific process of step S2 is:

[0014] Obtain pixel values ​​of a single-side fill-light image based on the Lambertian reflectance model The unit surface normal vector of the 2.5D image of the insulator surface corresponding to the positive projection of the corresponding fill light source The relationship between them is specifically:

[0015] ;

[0016] in is the insulator surface reflectivity, is the light source direction unit vector, is the light source intensity;

[0017] Then, based on the four non-collinear independent light fields, the normal vector set is obtained by least squares, specifically:

[0018] ;

[0019] in is the product of light source intensity and reflectivity, is the light source direction matrix, yes The transposed matrix of yes The inverse matrix of

[0020] Then obtain the unit normal vector set by normalization ;in Represents the surface unit normal vector in the x-axis, y-axis, and z-axis directions of space, and the elements Represented as the depth feature of each pixel.

[0021] Preferably, in step S3, singular value decomposition is performed on the 2.5D depth map of the insulator surface, and the singular value reconstructed image of the first part is retained to remove noise; and then multi-level Gaussian blurring is performed through Gaussian pyramid downsampling to generate a low-resolution image sequence that retains the overall structure.

[0022] Preferably, in step S4, in the process of obtaining the sample set A, a traditional bicubic B-spline surface fitting method is used to perform surface fitting;

[0023] In the process of obtaining the sample set B, an improved bicubic B-spline surface fitting method is used for surface fitting; in the improved bicubic B-spline surface fitting method, the denoised depth map is first gridded and sparsely processed to extract random control point coordinates.

[0024] Preferably, the specific process of obtaining the insulator surface curved model is:

[0025] Build based on The feature extraction network of sample set A and sample set B is trained through the network to obtain the final control point extractor , obtain the control points of the surface through GPU acceleration , and thus infer the surface model ;

[0026] in is the surface equation expression, and is the standard bicubic B-spline function; Then are the coefficient sets of the cubic spline function respectively; i 、 j Indicates the horizontal and vertical coordinate sequence numbers; m and n Indicates the number of horizontal and vertical control points.

[0027] Preferably, in step S5, the parameter information of the insulator surface includes the gradient distribution and the average curvature distribution of the insulator surface.

[0028] Preferably, in step S5, the variance of the observed surface average curvature distribution is compared with the expected surface average curvature distribution of the qualified sample; if the variance difference exceeds a preset threshold, it is determined that the insulator has surface deformation or defects.

[0029] The present invention also discloses an insulator surface defect detection system based on photometric stereo vision, comprising:

[0030] An image acquisition module, used for acquiring multiple insulator surface images;

[0031] A feature extraction module is used to restore a 2.5D depth map of the insulator surface from the insulator surface image based on photometric stereo.

[0032] An image denoising module is used to denoise the 2.5D depth map of the insulator surface to obtain a denoised depth map;

[0033] The surface model generation module is used to extract key points and perform surface fitting on the denoised depth map to obtain key control points and corresponding surfaces as sample set A; construct random control point coordinate combinations and corresponding surfaces as sample set B; and compare sample sets A and B based on the The feature extraction network is trained to obtain the final control point extractor to obtain the control points of the surface, and then the surface model of the insulator is obtained;

[0034] The defect detection module is used to calculate parameter information of the insulator surface based on the insulator surface curved model and perform defect detection according to the parameter information.

[0035] Preferably, the image acquisition module includes an optical unit and a lifting unit; the optical unit is located on the lifting unit; the optical unit includes a camera and a plurality of telecentric illumination light sources, and the plurality of telecentric illumination light sources are evenly distributed along the circumference of the camera.

[0036] Compared with the prior art, the advantages of the present invention are:

[0037] The present invention uses photometric stereo, multi-light source independent stroboscopic imaging, and a Lambertian reflectance model to accurately solve surface normal vectors and generate a 2.5D depth map with submillimeter resolution. This overcomes the shadow interference of monocular vision and the close-range stability issues of TOF cameras, significantly improving the detection capability of tiny defects (such as 0.1mm cracks). Based on bicubic B-spline surface fitting, the present invention utilizes local control characteristics to dynamically encrypt control points in defective areas, accurately reproducing local deformations (such as concave contours) under the premise of global smoothness, and avoiding the over-smoothing of details by traditional filtering algorithms. The gridded sparse sampling of the present invention reduces the number of control points to 1 / 100 to 1 / 50 of the original data. Combined with the GPU-accelerated HR-net key point detection algorithm, the control point extractor is trained, which increases the surface modeling speed by 3 to 5 times, meeting the needs of real-time detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 The figure is a flow chart of a defect detection method according to an embodiment of the present invention.

[0039] Figure 2 This is a front view of an embodiment of the defect detection system of the present invention.

[0040] Figure 3 A side view of an embodiment of a defect detection system of the present invention.

[0041] Figure 4 This is a diagram illustrating an embodiment of the defect detection method of the present invention in a specific application. DETAILED DESCRIPTION

[0042] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0043] like Figure 1 and Figure 4 As shown, the insulator surface defect detection method based on photometric stereo vision provided by the embodiment of the present invention specifically includes the following steps:

[0044] S1. Establish an optical environment for photometric stereo detection and collect multiple images of the insulator surface using photometric stereo.

[0045] The optical environment includes an optical unit and a lifting unit; the optical unit is located on the lifting unit; the optical unit includes a camera and multiple telecentric lighting sources, and the multiple telecentric lighting sources are evenly distributed around the camera. Specifically, with the lifting unit (such as a lifting robot) as the carrier, the optical unit follows the lifting robot's running path to cover the entire insulator surface for segmented inspection; the optical system includes a monocular CCD camera and four symmetrically distributed telecentric lighting sources (light sources 1#-4#), such as Figure 2-Figure 3 The camera is perpendicular to the robot's trajectory and forms an orthographic projection relationship. Each telecentric illumination light source is tilted at a 45° angle to the insulator surface, and the light source intensity is kept constant. , compared to the normal vector of the insulator surface is , which can be obtained by spherical calibration method; after establishing this optical environment, four local single-side images of the insulator are obtained by sequential stroboscopic photography using a monocular CCD camera. are the light source intensities of light sources 1#-4# respectively; They are the light source direction unit vectors of light sources 1#-4# respectively.

[0046] Of course, in other embodiments, 5, 6 or more light sources may be used and symmetrically distributed around the camera.

[0047] S2. Restore a 2.5D depth map of the insulator surface from the insulator surface image;

[0048] Based on the Lambertian reflectance model, a mapping relationship between the pixel values ​​of the single-side fill-light image and the surface normal vector is established. Combined with the light source direction matrix and light source intensity of the four groups of light sources, the normal vector set is solved using the least squares method and normalized to generate a 2.5D depth map matrix representing the surface depth.

[0049] Specifically, the pixel value of the single-side fill-in image can be obtained based on the Lambertian reflection principle: The unit surface normal vector of the 2.5D image of the insulator surface corresponding to the positive projection of the corresponding fill light source The relationship between them is specifically:

[0050] ;

[0051] in is the insulator surface reflectivity, is the light source direction unit vector, is the light source intensity;

[0052] Based on the four non-collinear independent light fields (the optical axes of the four light sources are not parallel to each other), the normal vector set can be obtained by the least squares method, specifically:

[0053] ;

[0054] in is the product of light source intensity and reflectivity, is the light source direction matrix, yes The transposed matrix of yes The inverse matrix of

[0055] Then obtain the unit normal vector set by normalization ;in Represents the surface unit normal vector in the x-axis, y-axis, and z-axis directions of space, and the elements Represented as the depth feature of each pixel.

[0056] S3. Denoising and smoothing of depth map

[0057] Perform singular value decomposition (SVD) on the depth map obtained in step S2, retaining the first 10% of singular values ​​to reconstruct the image to remove noise; then further downsample using a Gaussian pyramid and perform multi-level Gaussian blurring with a 5×5 Gaussian kernel to generate a low-resolution image sequence that retains the overall structure;

[0058] Specifically, due to It is a preliminary depth map matrix. Its uneven gradient will lead to too many noise points when extracting autocorrelation features. Therefore, singular value decomposition is used to decompose the depth map matrix. Denoising, specifically:

[0059] ;

[0060] in, and They are The orthogonal matrix corresponding to the length and width order, is the corresponding singular value diagonal matrix in descending order;

[0061] Pick The first 10% of the singular values ​​generate a diagonal matrix to reconstruct the image, which can best preserve local features while removing noise;

[0062] The reconstructed image is then subjected to pyramid downsampling. By alternating Gaussian blurring and downsampling, a series of images with gradually decreasing resolution are generated. Gradually reducing the resolution after Gaussian blurring preserves the overall image structure and minimizes detail loss. The number of pyramid layers should be limited to three; otherwise, the top layer's resolution will be too low, resulting in excessive information loss. A 5x5 Gaussian kernel is used, which minimizes computational effort and provides a good blurring effect.

[0063] S4. Surface Fitting and Model Construction Based on Improved B-Spline Interpolation

[0064] The denoised depth map is subjected to key point extraction and surface fitting to obtain key control points and corresponding surfaces as sample set A; random control point coordinate combinations and corresponding surfaces are constructed as sample set B; sample set A and sample set B are used to compare the key control points and corresponding surfaces based on the sample set A. The feature extraction network is trained to obtain the final control point extractor to obtain the control points of the surface, and then the surface model of the insulator is obtained;

[0065] Specifically, when calculating the control points of the denoised depth map, the traditional bicubic B-spline surface fitting method requires a large number of iterations, and the least squares method is used to fit the solution of the control points. The computing resources and speed need to be optimized, so grid data is used to perform sparse point cloud operations to reduce the amount of calculation; for example exist The mean value is extracted from the pixel values ​​to replace the depth information of the entire grid; Refers to the width and height of the grid value respectively. Usually a large enough grid is needed to ensure the smoothness of the surface and reduce the number of control points to the original indivual;

[0066] The random control point coordinate combination is constructed by the improved bicubic B-spline surface fitting method. And generate the corresponding surface and get the surface equation expression, specifically:

[0067] ;

[0068] in is the surface equation expression, and is the standard bicubic B-spline function; Then are the coefficient sets of the cubic spline function respectively; i 、 j Indicates the horizontal and vertical coordinate sequence numbers; m and n Indicates the number of horizontal and vertical control points; the control points constructed in this way and the corresponding surfaces are used as sample set B; the control point coordinate combination Include control points , It is a (n+1)×(m+1) grid that determines the surface shape.

[0069] For the denoised depth map, the traditional bicubic B-spline surface fitting method is used to obtain the control points and surfaces as the sample set A;

[0070] Build based on The feature extraction network is trained with sample sets A and B to obtain the final control point extractor. ,pass GPU Get the control points of the surface after acceleration , and thus the surface model is inferred:

[0071] ;

[0072] in Indicates smoothing or filtering of the node set N , and then according to the i-th sampling rule Generate control points. For example, use dense sampling in areas of high curvature , using sparse sampling in flat areas .

[0073] Here, B-spline-based surface fitting is used to obtain the insulator surface curve model, and then the original coordinate points are brought in to approximate the global appearance model while retaining the local features of the surface.

[0074] The present invention is based on surface fitting of bicubic B-splines and utilizes local control characteristics to dynamically encrypt control points in the defect area, accurately reproducing local deformation (such as concave contours) under the premise of global smoothness, avoiding excessive smoothing of details by traditional filtering algorithms.

[0075] The gridded sparse sampling of the present invention reduces the number of control points to 1 / 100-1 / 50 of the original data. Combined with the GPU-accelerated HR-net key point detection algorithm and training of the control point extractor, the surface modeling speed is increased by 3-5 times, meeting the real-time detection requirements.

[0076] S5. Surface defect detection and analysis

[0077] Based on the surface model obtained in step S4, the gradient distribution and average curvature distribution of the insulator surface are calculated, and the variance of the observed surface average curvature distribution is compared with the expected surface average curvature distribution of the qualified sample; if the variance difference exceeds a preset threshold, it is determined that the insulator has surface deformation or defects.

[0078] The relationship between the observed surface's gradient distribution and mean curvature distribution is essentially the inherent connection between the first-order differential properties of surface geometry (gradient) and the second-order differential properties (curvature), and can be used for surface morphology analysis. For example, the gradient distribution can be used to describe the surface's "slope," while the mean curvature distribution describes its "curvature."

[0079] Specifically, based on the inferred surface model , the depth value of any point on the insulator can be obtained, and then the gradient distribution and average curvature distribution can be calculated. By comparing this data with the data corresponding to qualified insulators, the quality of the insulator can be determined.

[0080] Specifically, the gradient calculation first uses numerical differentiation methods (such as central difference) to calculate the gradient component of each point, and then obtains the gradient modulus ;in represents the gradient component in the transverse direction, Represents the gradient component in the longitudinal direction; then, based on the gradient modulus, the following two judgment methods are used to determine whether the surface slope is qualified:

[0081] Global judgment: Count the maximum gradient modulus of the entire surface , whether it is lower than the inclination angle threshold T (the maximum allowable inclination is defined according to actual needs, such as T=0.21, corresponding to an inclination angle of about 12°). If it is lower than the threshold T, it means that the surface inclination is qualified, otherwise it is unqualified.

[0082] Local judgment: Mark all areas where the gradient modulus exceeds the threshold and analyze whether their area ratio is lower than the ratio threshold. ( Generally it is set to 5%). If the exceeding area does not exceed 5% of the total area, it means the surface slope is qualified, otherwise it is unqualified.

[0083] Specifically, for example, the expected average surface curvature distribution of an insulator is C1, and the observed average surface curvature distribution is C2. If the difference between the variance of C1 and the variance of C2 is greater than a certain threshold, it is considered that the insulator has surface deformation.

[0084] The present invention realizes defect determination based on autocorrelation characteristics (such as curvature variance comparison) without relying on a pre-annotated defect database, and is suitable for detecting unknown or rare defect types.

[0085] The embodiment of the present invention further provides an insulator surface defect detection system based on photometric stereo vision, comprising:

[0086] An image acquisition module, used for acquiring multiple insulator surface images;

[0087] A feature extraction module is used to restore a 2.5D depth map of the insulator surface from the insulator surface image based on photometric stereo.

[0088] An image denoising module is used to denoise the 2.5D depth map of the insulator surface to obtain a denoised depth map;

[0089] The surface model generation module is used to extract key points and perform surface fitting on the denoised depth map to obtain key control points and corresponding surfaces as sample set A; construct random control point coordinate combinations and corresponding surfaces as sample set B; and compare sample sets A and B based on the The feature extraction network is trained to obtain the final control point extractor to obtain the control points of the surface, and then the surface model of the insulator is obtained;

[0090] The defect detection module is used to calculate parameter information of the insulator surface based on the insulator surface curved model and perform defect detection according to the parameter information.

[0091] like Figure 2 and Figure 3 As shown, the image acquisition module includes an optical unit and a lifting unit; the optical unit is located on the lifting unit; the optical unit includes a camera and multiple telecentric illumination sources, which are evenly distributed around the camera. Specifically, a single camera plus four sets of surface light sources are fixed in a fixed position and tilted at a certain angle to form the optical system. The center point normals of the four sets of telecentric illumination sources converge on the central axis of the insulator surface and are distributed sequentially at 90° horizontal angles. The camera is located at the center of the light field and perpendicular to the insulator surface. The acquisition method is to flash a single surface light source sequentially as an independent light field. The camera captures four images of the single telecentric illumination source as the data basis for photometric stereo vision.

[0092] The four sets of telecentric light sources of the present invention are symmetrically distributed at a 45° tilt angle. Combined with the photometric stereo method, the surface normal vector is accurately solved through multi-light source independent stroboscopic imaging and the Lambertian reflection model to generate a 2.5D depth map with a sub-millimeter resolution, overcoming the shadow interference and TOF The camera's close-range stability significantly improves the ability to detect tiny defects (such as 0.1mm cracks).

[0093] An embodiment of the present invention further discloses a computer-readable storage medium having a computer program stored thereon. When executed by a processor, the computer program performs the steps of the above-described method. An embodiment of the present invention further discloses a computer device comprising a memory and a processor connected to each other. The memory has a computer program stored thereon. When executed by the processor, the computer program performs the steps of the above-described method. The medium and device of the present invention correspond to the above-described method and similarly possess the advantages of the above-described method.

[0094] The present invention can implement all or part of the process steps in the above-described method embodiments through hardware associated with computer program instructions. The computer program can be stored in a computer-readable storage medium. When executed by a processor, the computer program can implement the steps of the above-described method embodiments. The computer program includes computer program code, which can be in source code form, object code form, executable file, or some intermediate form. Computer-readable storage media include any entity or device capable of carrying computer program code, recording media, USB flash drives, removable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media. Memory is used to store computer programs and / or modules. The processor implements various functions by running or executing the computer programs and / or modules stored in the memory and accessing data stored in the memory. The memory may include a high-speed random access memory and may also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a Smart Media Card (SMC), a Secure Digital (SD) card, a Flash Card, at least one disk storage device, a flash memory device, or other volatile solid-state storage device.

[0095] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions based on the principles of the present invention are within the scope of protection of the present invention. It should be noted that for those skilled in the art, various improvements and modifications that do not depart from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A method for detecting surface defects of insulators based on photometric stereo vision, characterized in that: Including steps: S1. Acquire multiple insulator surface images obtained by photometric stereo; S2. Restore a 2.5D depth map of the insulator surface from the insulator surface image; S3. Denoise the 2.5D depth map of the insulator surface to obtain a denoised depth map; S4. Extract key points and perform surface fitting on the denoised depth map to obtain key control points and corresponding surfaces as sample set A; construct random control point coordinate combinations and corresponding surfaces as sample set B; Through the sample set A and sample set B based on The feature extraction network is trained to obtain the final control point extractor to obtain the control points of the surface, and then the surface model of the insulator is obtained; S5. Calculate parameter information of the insulator surface based on the insulator surface model and perform defect detection based on the parameter information; In step S2, a mapping relationship is established between the pixel values ​​of the single-sided fill light image and the unit surface normal vector of the 2.5D image of the insulator surface corresponding to the orthographic projection of the corresponding fill light source based on the Lambertian reflectance model. The light source direction matrix and light source intensity of multiple groups of light sources are combined to solve the normal vector set by the least squares method, and the normal vector set is normalized to generate a 2.5D depth map representing the surface depth, that is, the depth map matrix; The specific process of step S2 is: Obtain pixel values ​​of a single-side fill-light image based on the Lambertian reflectance model The unit surface normal vector of the 2.5D image of the insulator surface corresponding to the orthographic projection of the corresponding fill light source The relationship between them is specifically: ； in is the insulator surface reflectivity, is the light source direction unit vector, is the light source intensity; Then, based on the four non-collinear independent light fields, the normal vector set is obtained by least squares, specifically: ； in is the product of light source intensity and reflectivity, is the light source direction matrix, yes The transposed matrix of yes The inverse matrix of Then obtain the unit normal vector set by normalization ;in Represents the surface unit normal vector in the x-axis, y-axis, and z-axis directions of space, and the elements Represented as the depth feature of each pixel.

2. The insulator surface defect detection method based on photometric stereo vision according to claim 1, characterized in that: In step S3, singular value decomposition is performed on the 2.5D depth map of the insulator surface, and the singular value reconstructed image of the first part is retained to remove noise; then, Gaussian pyramid downsampling is further performed, and multi-level Gaussian blur is performed to generate a low-resolution image sequence that retains the overall structure.

3. The insulator surface defect detection method based on photometric stereo vision according to claim 1, characterized in that: In step S4, in the process of obtaining the sample set A, the traditional bicubic B-spline surface fitting method is used to perform surface fitting; In the process of obtaining the sample set B, an improved bicubic B-spline surface fitting method is used for surface fitting; in the improved bicubic B-spline surface fitting method, the denoised depth map is first gridded and sparsely processed to extract random control point coordinates.

4. The insulator surface defect detection method based on photometric stereo vision according to claim 3, characterized in that: The specific process of obtaining the insulator surface model is as follows: Build based on The feature extraction network of sample set A and sample set B is trained through the network to obtain the final control point extractor , obtain the control points of the surface through GPU acceleration , and thus infer the surface model ; in is the surface equation expression, and is the standard bicubic B-spline function; Then are the coefficient sets of the cubic spline function respectively; i 、 j Indicates the horizontal and vertical coordinate sequence numbers; m and n Indicates the number of horizontal and vertical control points.

5. The insulator surface defect detection method based on photometric stereo vision according to claim 1, 2 or 3, characterized in that: In step S5 , the parameter information of the insulator surface includes the gradient distribution and the average curvature distribution of the insulator surface.

6. The insulator surface defect detection method based on photometric stereo vision according to claim 5, characterized in that: In step S5, the variance of the observed surface average curvature distribution is compared with the expected surface average curvature distribution of the qualified sample; If the variance difference exceeds a preset threshold, it is determined that the insulator has surface deformation or defects.

7. An insulator surface defect detection system based on photometric stereo vision, used to execute the steps of the insulator surface defect detection method based on photometric stereo vision according to any one of claims 1 to 6, characterized in that: The detection system includes: An image acquisition module, used for acquiring multiple insulator surface images; A feature extraction module is used to restore a 2.5D depth map of the insulator surface from the insulator surface image based on photometric stereo. An image denoising module is used to denoise the 2.5D depth map of the insulator surface to obtain a denoised depth map; The surface model generation module is used to extract key points and perform surface fitting on the denoised depth map to obtain key control points and corresponding surfaces as sample set A; construct random control point coordinate combinations and corresponding surfaces as sample set B; and compare sample sets A and B based on the The feature extraction network is trained to obtain the final control point extractor to obtain the control points of the surface, and then the surface model of the insulator is obtained; The defect detection module is used to calculate parameter information of the insulator surface based on the insulator surface curved model and perform defect detection according to the parameter information.

8. The insulator surface defect detection system based on photometric stereo vision according to claim 7, characterized in that: The image acquisition module includes an optical unit and a lifting unit; the optical unit is located on the lifting unit; the optical unit includes a camera and a plurality of telecentric illumination light sources, and the plurality of telecentric illumination light sources are evenly distributed along the circumference of the camera.

Images