Laser scribing detection method for solar cell
Through the frequency domain feature model and region recognition model of the deep learning model, combined with preprocessing and fusion technology, the problems of stability and accuracy of solar cell laser scribing detection in complex environments are solved, and efficient automatic laser scribing recognition is achieved.
Patent Information
- Application Number
- CN202510188714.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-20
- Publication Date
- 2025-06-10
AI Technical Summary
The laser scribe detection of solar cells is difficult to maintain stability in complex visual environments, and uneven coating and material accumulation affect the detection accuracy.
The frequency domain feature model and region recognition model of the deep learning model are adopted. By preprocessing and fusion of the target image, frequency domain features and spatial domain features are extracted, and combined with the frequency domain weight layer and feature processing module, the scribing area is automatically identified and the laser scribing is found.
It improves the accuracy and stability of laser scribing detection, reduces the need for manual adjustment, adapts to various coated surface features in complex process environments, and improves the accuracy and robustness of image recognition.
Smart Images
Figure CN120125528A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the fields of laser cutting and etching, and particularly to a laser scribing detection method for solar cells. Background Art
[0002] In the laser scribing process of solar cells, strict control of the dead zone is required to increase the power generation area of the cells. This requires that the line spacing between the scribes P1 / P2 / P3 on the solar cell be as small as possible. If the scribes between P1 / P2 / P3 are not parallel, in order to avoid scribe crossing, the scribe spacing needs to be increased. Since P1 / P2 / P3 are the upstream and downstream process connection devices of the same product on the production line, that is, after P1 scribing, it is necessary to reposition the scribe P1 on the device of P2. Similarly, for P3 scribing, it is also necessary to reposition P1 or P2 scribing.
[0003] After positioning the above scribes, it is necessary to perform angle compensation on the current scribes in real time. Common angle compensation methods mainly perform mechanical motion interpolation of the axis through the servo system or correct the position of the glass plate through the machine vision system.
[0004] However, the solar cell process is currently in the exploratory stage, and the process preparation ability is not yet stable. Therefore, uneven coating thickness, coating defects, and coating liquid accumulation on the surface of the prepared glass plate will reduce the detection accuracy of the vision algorithm. And the recognition ability of the conventional vision algorithm is limited, and it is difficult to remain stable in a complex vision environment, often requiring manual adjustment of vision parameters, which reduces the production efficiency and production quality. Summary of the Invention
[0005] The purpose of this application is to provide a laser scribing detection method for solar cells, which can improve the scribing detection accuracy and remain stable in a complex vision environment.
[0006] To achieve the above purpose, this application provides a laser scribing detection method for solar cells, including:
[0007] Preprocessing the target image to obtain a first spatial domain image;
[0008] Obtaining a corresponding frequency domain image according to the first spatial domain image;
[0009] After preprocessing the frequency domain image, obtaining a corresponding second spatial domain image according to the frequency domain image;
[0010] Fusing the first spatial domain image and the second spatial domain image to obtain a fused image;
[0011] Inputting the frequency domain image into a frequency domain feature model to obtain frequency domain features, and the frequency domain feature model is a deep learning model;
[0012] Input the frequency domain features and the fused image into a region recognition model to obtain a marked region, where the region recognition model is a deep learning model;
[0013] Find the laser marking within the marked region.
[0014] Optionally, the formula for obtaining the corresponding frequency domain image from the first spatial domain image includes:
[0015]
[0016] where x and y are the pixel coordinates of the first spatial domain image, u and v are the frequency coordinates of the frequency domain image, f(x, y) is the pixel value of the first spatial domain image, F(u, v) is the complex value of the frequency domain image, M and N are the width and height of the first spatial domain image respectively, and j is the imaginary unit, j 2 = -1.
[0017] Optionally, the formula for preprocessing the frequency domain image includes:
[0018] G(u, v) = H(u, v) * F(u, v);
[0019] where u and v are the frequency coordinates of the frequency domain image, F(u, v) is the value of the frequency domain image, H(u, v) is an image filter, and G(u, v) is the corresponding pixel value after preprocessing.
[0020] Optionally, H(u, v) is a high-pass filter in the frequency domain, and the formula for H(u, v) includes:
[0021]
[0022] where D 0 is the cut-off frequency.
[0023] Optionally, the formula for fusing the first spatial domain image and the second spatial domain image to obtain a fused image includes:
[0024]
[0025] where x and y are the pixel coordinates of the first spatial domain image, u and v are the frequency coordinates of the frequency domain image, K is a fusion coefficient, f(x, y) is the pixel value of the first spatial domain image, G(u, v) is the pixel value after preprocessing the frequency domain image, M and N are the width and height of the first spatial domain image respectively, and j is the imaginary unit, j 2 = -1, and f'(x, y) is the pixel value of the fused image.
[0026] Optionally, the frequency domain feature model is a graph convolutional model;
[0027] The frequency domain feature model is used to perform inverse graph Fourier transform convolution on the frequency domain image to obtain the frequency domain features.
[0028] Optionally, the region recognition model includes a frequency domain weight layer and a feature processing module;
[0029] The step of inputting the frequency domain features and the fused image into the region recognition model to obtain the marked region includes:
[0030] Performing feature extraction on the frequency domain features to obtain a frequency domain feature tensor;
[0031] Performing feature extraction on the fused image to obtain a spatial domain feature tensor;
[0032] Concatenating the frequency domain feature tensor and the spatial domain feature tensor to obtain a fused tensor;
[0033] Inputting the fused tensor into the frequency domain weight layer, and the frequency domain weight layer adjusts the weight of the frequency domain feature tensor in the fused tensor;
[0034] Using the feature processing module to obtain the marked region according to the output result of the frequency domain weight layer.
[0035] Optionally, the calculation formula of the frequency domain weight layer includes:
[0036] W freq = γ * X m-1 ;
[0037] where γ is a learnable parameter value, X m-1 is the frequency domain feature tensor in the fused tensor, and W freq is the frequency domain feature tensor with adjusted weight.
[0038] Optionally, finding the laser marking line within the marked region includes:
[0039] Obtaining a marked grayscale image according to the image corresponding to the marked region;
[0040] On the marked grayscale image, performing edge feature enhancement along the direction of the laser marking line using a differential template;
[0041] Performing threshold segmentation on the marked grayscale image;
[0042] Performing grayscale projection in the direction of the laser marking line on the marked grayscale image to obtain a grayscale projection function;
[0043] Take the first derivative of the grayscale projection function in the direction perpendicular to the laser scribing to obtain the grayscale change function, and find the local extrema of the grayscale change function to obtain the grayscale maximum point and the grayscale minimum point on the scribing grayscale image;
[0044] Perform local grayscale value change detection on the pixel points between the grayscale maximum point and the grayscale minimum point to obtain a number of edge points;
[0045] Fit a straight line using each of the edge points to obtain the laser scribing.
[0046] Optionally, before performing edge feature enhancement using a difference template in the direction along the laser scribing, it further includes:
[0047] Smooth and filter the scribing grayscale image.
[0048] Optionally, if the direction of the laser scribing is roughly along the x - direction of the scribing grayscale image, the formula for performing edge feature enhancement using a difference template in the direction along the laser scribing includes:
[0049]
[0050] where G x (m,n) is the difference template in the x - direction, I is the scribing grayscale image, and G x is the pixel value after edge feature enhancement of the pixel points of the scribing grayscale image.
[0051] Optionally, if the direction of the laser scribing is roughly along the y - direction of the scribing grayscale image, the formula for performing feature point enhancement using a difference template in the direction along the laser scribing includes:
[0052]
[0053] where G y (m,n) is the difference template in the y - direction, I is the scribing grayscale image, and G y is the pixel value after edge feature enhancement of the pixel points of the scribing grayscale image.
[0054] Optionally, the threshold segmentation of the scribing grayscale image includes:
[0055] Use the watershed algorithm to determine the overexposed part and the effective part of the scribing grayscale image;
[0056] Approximate the grayscale value of each pixel point in the overexposed part to the mean value of the grayscale values of each pixel point in the effective part;
[0057] Use the maximum inter - class variance method to perform threshold segmentation on the scribing grayscale image.
[0058] Optionally, the formula for approximating the gray value of each pixel point in the overexposed part to the average value of the gray values of each pixel point in the valid part includes:
[0059] G′ (x,y) = G (x,y) - α * (G (x,y) - G effective );
[0060] Wherein, G (x,y) is the gray value corresponding to the pixel point (x, y) in the overexposed part, α is a constant, G effective is the average value of the gray values of each pixel point in the valid part, and G′ (x,y) is the adjusted gray value of the pixel point with coordinates (x, y).
[0061] Optionally, performing local gray value change detection on the pixel points between the gray value maximum point and the gray value minimum point to obtain a plurality of edge points includes:
[0062] Performing local gray value change detection from the gray value maximum point to the gray value minimum point, and when it is detected that the gray value of a pixel point starts to increase or decrease, determining that pixel point as the edge point;
[0063] Optionally, performing local gray value change detection from the gray value minimum point to the gray value maximum point, and when it is detected that the gray value of a pixel point starts to increase or decrease, determining that pixel point as the edge point.
[0064] Compared with the prior art, the present application performs data enhancement on a target image to obtain a first spatial domain image; obtains a corresponding frequency domain image according to the first spatial domain image; after preprocessing the frequency domain image, obtains a corresponding second spatial domain image according to the frequency domain image; fuses the first spatial domain image and the second spatial domain image to obtain a fused image; inputs the frequency domain image into a frequency domain feature model to obtain frequency domain features, inputs the frequency domain features and the fused image into a region recognition model to obtain a marked region, and finds a laser mark within the marked region, which can reduce the interference caused by the uneven thickness, coating defects, and coating liquid accumulation that easily occur on the coated surface of the glass plate of a solar cell, improve the mark detection accuracy, and maintain stability in a complex visual environment. In addition, since the frequency domain feature model and the region recognition model are deep learning models, compared with traditional rule-based image processing methods, the laser mark detection method of the present application can adapt to various types of coated surface features through training in a complex process environment, thereby improving the accuracy and robustness of image recognition. BRIEF DESCRIPTION OF THE DRAWINGS
[0065] Figure 1 is a flowchart of the laser mark detection method according to an embodiment of the present application.
[0066] Figure 2 This is the principle block diagram of the laser scribing detection method according to the embodiments of the present application.
[0067] Figure 3 、 Figure 4 and Figure 5 are respectively other flowcharts of the laser scribing detection method according to the embodiments of the present application. Detailed implementation manners
[0068] To describe in detail the technical content, structural features, achieved objectives and effects of the present application, the following will be described in detail in conjunction with the embodiments and with reference to the drawings.
[0069] In the laser scribing process of perovskite solar cells, it is necessary to accurately position the laser scribing and control the scribing spacing well, especially to perform visual error compensation in complex process environments such as uneven surface coating, coating defects and material accumulation. Traditional vision algorithms are difficult to stably locate laser scribing in these complex environments and require frequent manual adjustment, which affects production efficiency and quality. Therefore, a laser scribing detection method is needed that can achieve high-precision automatic recognition of the scribing position, angle and spacing on the glass plate surface, improve production efficiency while reducing manual intervention, and ensure the scribing accuracy and spacing control between P1 / P2 / P3 devices, thereby maximizing the power generation area and conversion efficiency of the solar cells.
[0070] The frequency domain feature model and region recognition model involved in the present application are deep learning models. Among them, the original data set for training the frequency domain feature model and region recognition model can be obtained by collecting backlight images of glass plates with various coating abnormal states, different light source brightnesses, exposure brightnesses and containing P1, P2 and P3 scribings, so as to obtain the original data set. During training, target images can be obtained from the original data set, and the target images can be preprocessed by methods such as mixed color jitter, local occlusion, affine transformation and Gaussian noise, so as to expand and enhance the original data set and improve the quality of the image data in the original data set. It can be understood that when actually using the frequency domain feature model and region recognition model for laser scribing detection, the preprocessing method of the target image is not necessarily the same as that during training, which is a well-known technical means in the art and will not be elaborated here.
[0071] Please refer to Figure 1 and Figure 2 , the present application discloses a laser scribing detection method for solar cells, including:
[0072] S1, preprocess the target image to obtain a first spatial domain image.
[0073] S2, obtain the corresponding frequency domain image according to the first spatial domain image.
[0074] S3. After preprocessing the frequency-domain image, obtain the corresponding second spatial-domain image according to the frequency-domain image.
[0075] S4. Fuse the first spatial-domain image and the second spatial-domain image to obtain a fused image.
[0076] S5. Input the frequency-domain image into the frequency-domain feature model 1 to obtain frequency-domain features.
[0077] S6. Input the frequency-domain features and the fused image into the region recognition model 2 to obtain the marked region.
[0078] S7. Locate the laser marking within the marked region.
[0079] It can be understood that there is no sequential order between steps S3 - S4 and step S5. Step S5 can also be executed first, and then steps S3 - S4.
[0080] Compared with the prior art, in this application, data augmentation is performed on the target image to obtain the first spatial-domain image; the corresponding frequency-domain image is obtained according to the first spatial-domain image; after preprocessing the frequency-domain image, the corresponding second spatial-domain image is obtained according to the frequency-domain image; the first spatial-domain image and the second spatial-domain image are fused to obtain a fused image; the frequency-domain image is input into the frequency-domain feature model 1 to obtain frequency-domain features, the frequency-domain features and the fused image are input into the region recognition model 2 to obtain the marked region, and the laser marking is located within the marked region, which can reduce the interference caused by the uneven thickness, coating defects, coating liquid accumulation, etc. that easily occur on the coated surface of the glass plate of the solar cell, improve the marking detection accuracy, and maintain stability in a complex visual environment. In addition, since the frequency-domain feature model 1 and the region recognition model 2 are deep learning models, compared with traditional rule-based image processing methods, the laser marking detection method of this application can adapt to various types of coated surface features through training in a complex process environment, thereby improving the accuracy and robustness of image recognition.
[0081] In some embodiments, a high-resolution camera can be used to continuously scan the surface of the glass plate to obtain the target image in real time.
[0082] In the frequency domain, the coated background with a dark background and the laser marking with a bright foreground have different characteristics, and the result in the frequency domain is enhanced through the frequency-domain feature enhancement formula.
[0083] In some embodiments, the formula for obtaining the corresponding frequency-domain image according to the first spatial-domain image includes:
[0084]
[0085] Among them, x and y are the pixel coordinates of the first spatial domain image, u and v are the frequency coordinates of the frequency domain image, f(x, y) is the pixel value of the first spatial domain image, F(u, v) is the complex value of the frequency domain image, M and N are the width and height of the first spatial domain image respectively, and j is the imaginary unit, where j 2 = -1. This formula is obtained based on the Fourier transform formula.
[0086] In some embodiments, the formula for preprocessing the frequency domain image includes:
[0087] G(u, v) = H(u, v) * F(u, v);
[0088] Among them, u and v are the frequency coordinates of the frequency domain image, F(u, v) is the value on the frequency domain image, H(u, v) is the image filter, and G(u, v) is the corresponding value after preprocessing.
[0089] Specifically, H(u, v) is a high-pass filter in the frequency domain, and the formula of H(u, v) includes:
[0090]
[0091] Among them, D 0 is the cut-off frequency, and u and v are the frequency coordinates in the frequency domain.
[0092] In some embodiments, the formula for fusing the first spatial domain image and the second spatial domain image to obtain a fused image includes:
[0093]
[0094] Among them, x and y are the pixel coordinates of the first spatial domain image, u and v are the frequency coordinates of the frequency domain image, K is the fusion coefficient, which is a constant rational number, f(x, y) is the pixel value of the first spatial domain image, G(u, v) is the pixel value after preprocessing the frequency domain image, M and N are the width and height of the first spatial domain image respectively, j is the imaginary unit, where j 2 = -1, and f'(x, y) is the pixel value of the fused image.
[0095] Specifically, the corresponding second spatial domain image can be obtained from the frequency domain image by performing an inverse discrete Fourier transform on the frequency domain image.
[0096] In some embodiments, the frequency domain feature model 1 is a graph convolutional model (GNN, Graph Neural Network). The frequency domain feature model 1 is used to perform an inverse graph Fourier transform convolution on the frequency domain image to obtain frequency domain features.
[0097] Please refer to Figure 2 and Figure 3, in some embodiments, the region recognition model 2 includes a frequency domain weight layer 21 and a feature processing module 22.
[0098] Inputting the frequency domain features and the fused image into the region recognition model 2 to obtain the marked area includes:
[0099] S61, performing feature extraction on the frequency domain features to obtain a frequency domain feature tensor.
[0100] S62, performing feature extraction on the fused image to obtain a spatial domain feature tensor.
[0101] Optionally, a convolutional layer (such as a CNN network, Convolutional Neural Networks) is used to perform feature extraction on the fused image to obtain a spatial domain feature tensor.
[0102] S63, concatenating (Concat) the frequency domain feature tensor and the spatial domain feature tensor to obtain a fused tensor.
[0103] Optionally, before concatenating the frequency domain feature tensor and the spatial domain feature tensor to obtain a fused tensor, it further includes adjusting the dimensions of the frequency domain feature tensor and / or the spatial domain feature tensor so that the dimensions of the frequency domain feature tensor and the spatial domain feature tensor are consistent and can be concatenated. It can be understood that if the dimensions of the frequency domain feature tensor and the spatial domain feature tensor are consistent, no adjustment is required.
[0104] S64, inputting the fused tensor into the frequency domain weight layer 21, and the frequency domain weight layer 21 adjusts the weight of the frequency domain feature tensor in the fused tensor.
[0105] Specifically, the calculation formula of the frequency domain weight layer 21 includes:
[0106] W freq = γ * X m-1 ;
[0107] where γ is a learnable parameter value, X m-1 is the frequency domain feature tensor in the fused tensor, and W freq is the frequency domain feature tensor with adjusted weight.
[0108] Optionally, after adjusting the weight of the frequency domain feature tensor through the above calculation formula, convolutional feature extraction is performed on the fused tensor, and the frequency domain weight layer 21 outputs the fused tensor after convolutional feature extraction.
[0109] S65, using the feature processing module 22 to obtain the marked area according to the output result of the frequency domain weight layer 21.
[0110] Please refer to Figure 2, specifically, the feature processing module 22 of the region recognition model 2 includes a residual convolutional layer 221, a region proposal network 222 (Region Proposal Network, RPN), a pooling layer 223 (Pooling Layer), and a fully connected layer 224 (Fully Connected Layer) connected in sequence. After the output result of the frequency domain weight layer 21 is output to the residual convolutional layer 221, the marked area is output by the fully connected layer 224. More specifically, the region recognition model 2 can be improved from the Faster RCNN model by adding the frequency domain weight layer 21. Of course, it is not limited to using the Faster RCNN model.
[0111] Please refer to Figure 4 , in some embodiments, finding the laser scribing line within the marked area includes:
[0112] S71, obtaining a scribing grayscale image according to the image corresponding to the marked area.
[0113] S72, on the scribing grayscale image, enhancing the edge features using a difference template along the direction of the laser scribing line.
[0114] S73, performing threshold segmentation on the scribing grayscale image.
[0115] S74, performing a grayscale projection in the direction of the laser scribing line on the scribing grayscale image to obtain a grayscale projection function;
[0116] S75, taking the first derivative of the grayscale projection function along the direction perpendicular to the laser scribing line to obtain a grayscale change function, and finding the local extreme values of the grayscale change function to obtain the grayscale maximum point and the grayscale minimum point on the scribing grayscale image.
[0117] S76, performing local grayscale value change detection on the pixel points between the grayscale maximum point and the grayscale minimum point to obtain a number of edge points.
[0118] S77, fitting a straight line using each edge point to obtain the laser scribing line.
[0119] Specifically, the least squares method can be used to fit a straight line using each edge point to obtain the laser scribing line. Of course, it is not limited to this.
[0120] Specifically, before enhancing the edge features using a difference template along the direction of the laser scribing line, it further includes:
[0121] Smoothing and filtering the scribing grayscale image. More specifically, some edge burrs and coating defects on the scribing grayscale image are removed through Gaussian filtering and image smoothing.
[0122] During laser scribing, the direction of the laser scribing line approximately coincides with the x direction or the y direction of the scribing grayscale image.
[0123] Specifically, if the direction of laser scribing is generally along the x-direction of the scribing grayscale image, the formula for edge feature enhancement using a difference template along the direction of laser scribing includes:
[0124]
[0125] Among them, G x (m,n) is the difference template in the x-direction, I is the scribing grayscale image, and G x is the pixel value after edge feature enhancement of the pixel points of the scribing grayscale image.
[0126] Specifically, if the direction of laser scribing is generally along the y-direction of the scribing grayscale image, the formula for feature point enhancement using a difference template along the direction of laser scribing includes:
[0127]
[0128] Among them, G y (m,n) is the difference template in the y-direction, i is the scribing grayscale image, and G y is the pixel value after edge feature enhancement of the pixel points of the scribing grayscale image.
[0129] Among them, the essence of calculating using the difference template is to calculate the gradient of the scribing grayscale image in the direction of the laser scribing line. After obtaining the gradient, non-maximum suppression is performed according to the gradient magnitude and direction, only retaining the points with the largest gradient value and smoothing the points with smaller gradient values around. Remove the thinning noise points at the edge of the laser scribing line to achieve the effect of enhancing the feature points of the edge.
[0130] Since the problems of uneven coating and material accumulation will seriously affect the edge recognition, the conventional threshold segmentation cannot effectively segment the coating liquid background and the scribing line foreground in the scribing grayscale image. In this regard, the present application performs threshold segmentation on the image in multiple steps through a cross-region threshold segmentation method to extract the effective foreground in the image and remove the overexposed and target-lost background part in the image.
[0131] Please refer to Figure 5 , specifically, the threshold segmentation of the scribing grayscale image includes:
[0132] S731, use the watershed algorithm to determine the overexposed part and the effective part of the scribing grayscale image.
[0133] S732, approximate the grayscale value of each pixel point in the overexposed part to the mean value of the grayscale values of each pixel point in the effective part.
[0134] More specifically, the formula for approximating the grayscale value of each pixel point in the overexposed part to the mean value of the grayscale values of each pixel point in the effective part includes:
[0135] G' (x,y) = G (x,y) - α * (G (x,y) - G effective );
[0136] Wherein, G (x,y) is the gray value corresponding to the pixel point (x, y) in the overexposed part, α is a constant, and G effective is the average value of the gray values of each pixel point in the valid part, and G' (x,y) is the adjusted gray value of the pixel point with coordinates (x, y). By adjusting the value of α, the approximation degree of the overexposed part to the valid part can be adjusted.
[0137] S733, perform threshold segmentation on the image using the maximum inter-class variance method.
[0138] Specifically, perform local gray value change detection on the pixel points between the maximum gray point and the minimum gray point to obtain several edge points, including:
[0139] Perform local gray value change detection from the maximum gray point to the minimum gray point. When it is detected that the gray value of the pixel point starts to increase or decrease, determine that pixel point as an edge point. And / or, perform local gray value change detection from the minimum gray point to the maximum gray point. When it is detected that the gray value of the pixel point starts to increase or decrease, determine that pixel point as an edge point.
[0140] Wherein, for local gray value change detection, when the gray value of the pixel point starts to increase, it may indicate the start of an object boundary, and when the gray value of the pixel point starts to increase, it may indicate the end of an object boundary.
[0141] Optionally, performing local gray value change detection from the maximum gray point to the minimum gray point and performing local gray value change detection from the minimum gray point to the maximum gray point can improve the detection accuracy.
[0142] The above-disclosed are only the preferred embodiments of the present application. Of course, the scope of rights of the present application cannot be limited thereby. Therefore, equivalent changes made according to the patent scope of the present application still fall within the scope covered by the present application.
Claims
1. A laser scribing detection method for solar cells, characterized in that: include: Preprocessing the target image to obtain a first spatial domain image; Obtaining a corresponding frequency domain image according to the first spatial domain image; After preprocessing the frequency domain image, obtaining a corresponding second spatial domain image according to the frequency domain image; fusing the first spatial domain image and the second spatial domain image to obtain a fused image; Inputting the frequency domain image into a frequency domain feature model to obtain frequency domain features, wherein the frequency domain feature model is a deep learning model; Inputting the frequency domain features and the fused image into a region recognition model to obtain a lined region, wherein the region recognition model is a deep learning model; The laser scribe line is found within the scribe area.
2. The laser scribing detection method according to claim 1, characterized in that: The formula for obtaining the corresponding frequency domain image according to the first spatial domain image includes: Wherein, x and y are the pixel coordinates of the first spatial domain image, u and v are the frequency coordinates of the frequency domain image, f(x, y) is the pixel value of the first spatial domain image, F(u, v) is the complex value of the frequency domain image, M and N are the width and height of the first spatial domain image, respectively, j is an imaginary unit, j 2 =-1.
3. The laser scribing detection method according to claim 1, characterized in that: The formula for preprocessing the frequency domain image includes: G(u,v)=H(u,v)*F(u,v); Among them, u and v are the frequency coordinates of the frequency domain image, F(u,v) is the value of the frequency domain image, H(u,v) is the image filter, and G(u,v) is the corresponding pixel value after preprocessing. Wherein, H(u,v) is a high-pass filter in the frequency domain, and the formula of H(u,v) includes: Where D0 is the cut-off frequency.
4. The laser scribing detection method according to claim 1, characterized in that: The formula for fusing the first spatial domain image and the second spatial domain image to obtain a fused image includes: Wherein, x and y are the pixel coordinates of the first spatial domain image, u and v are the frequency coordinates of the frequency domain image, K is the fusion coefficient, f(x, y) is the pixel value of the first spatial domain image, G(u, v) is the pixel value of the frequency domain image after preprocessing, M and N are the width and height of the first spatial domain image, j is the imaginary unit, j 2 =-1, f ′ (x, y) is the pixel value of the fused image.
5. The laser scribing detection method according to claim 1, characterized in that: The frequency domain feature model is a graph convolution model; The frequency domain feature model is used to perform inverse Fourier transform convolution on the frequency domain image to obtain the frequency domain features.
6. The laser scribing detection method according to claim 1, characterized in that: The region recognition model includes a frequency domain weight layer and a feature processing module; The step of inputting the frequency domain feature and the fused image into a region recognition model to obtain a lined region comprises: Extracting the frequency domain features to obtain a frequency domain feature tensor; Performing feature extraction on the fused image to obtain a spatial domain feature tensor; Concatenating the frequency domain feature tensor and the space domain feature tensor to obtain a fused tensor; Inputting the fused tensor into the frequency domain weight layer, and adjusting the weight of the frequency domain feature tensor in the fused tensor by the frequency domain weight layer; The feature processing module is used to obtain the marked area according to the output result of the frequency domain weight layer.
7. The laser scribing detection method according to claim 6, characterized in that: The calculation formula of the frequency domain weight layer includes: W freq =γ*X m-1 ; Among them, γ is the learnable parameter value, X m-1 is the frequency domain feature tensor in the fusion tensor, W jreq is the frequency domain feature tensor after adjusting the weight.
8. The laser scribing detection method according to claim 1, characterized in that: Finding a laser scribe line within the scribe area includes: Obtaining a line grayscale image according to the image corresponding to the line area; On the scribing grayscale image, edge feature enhancement is performed along the direction of the laser scribing using a differential template; Performing threshold segmentation on the line grayscale image; Performing grayscale projection on the scribing grayscale image in the laser scribing direction to obtain a grayscale projection function; Taking the first-order derivative of the grayscale projection function along a direction perpendicular to the laser scribing line to obtain a grayscale change function, and taking the local extreme value of the grayscale change function to obtain the grayscale maximum point and the grayscale minimum point on the scribing grayscale map; Performing local grayscale value change detection on pixel points between the grayscale maximum point and the grayscale minimum point to obtain a plurality of edge points; The laser scribing is obtained by fitting a straight line using each of the edge points. Before the edge feature enhancement is performed using a differential template along the direction of laser scribing, the method further includes: The line grayscale image is smoothed and filtered. Wherein, if the direction of the laser scribing is substantially along the x direction of the scribing grayscale image, the formula for edge feature enhancement using a differential template along the direction of the laser scribing includes: Among them, G x (m,n) is the differential template in the x direction, I is the grayscale image of the line, G x The pixel value of the pixel point of the line grayscale image after edge feature enhancement. and / or, If the direction of the laser scribing is roughly along the y direction of the scribing grayscale image, the formula for enhancing the feature points using the differential template along the direction of the laser scribing includes: Among them, G y (m,n) is the differential template in the y direction, I is the grayscale image of the line, G y The pixel value of the pixel point of the line grayscale image after edge feature enhancement.
9. The laser scribing detection method according to claim 8, characterized in that: The threshold segmentation of the line grayscale image comprises: Determine the overexposed part and the valid part of the line grayscale image by using a watershed algorithm; Approaching the grayscale value of each pixel in the overexposed part to the average grayscale value of each pixel in the effective part; The maximum inter-class variance method is used to perform threshold segmentation on the line grayscale image. The formula for approximating the grayscale value of each pixel in the overexposed part to the average grayscale value of each pixel in the effective part includes: G′ (x,y) =G (x,y) -a*(G (x,y) -G effective ); Among them, G (x,y) is the grayscale value corresponding to the pixel point (x, y) of the overexposed part, α is a constant, G effective is the mean gray value of each pixel in the effective part, G′ (x,y) The grayscale value after adjustment for the pixel point (x, y).
10. The laser scribing detection method according to claim 8, characterized in that: The performing of local grayscale value change detection on the pixel points between the grayscale maximum point and the grayscale minimum point to obtain a plurality of edge points includes: Performing local grayscale change detection from the grayscale maximum point to the grayscale minimum point, and when it is detected that the grayscale value of a pixel point begins to increase or decrease, determining the pixel point as the edge point; and / or, A local grayscale change detection is performed from the grayscale minimum point to the grayscale maximum point, and when it is detected that the grayscale value of a pixel point begins to increase or decrease, the pixel point is determined to be the edge point.