Solar cell defect detection system based on machine vision and deep learning
By decomposing a two-dimensional image into a one-dimensional sequence and constructing a protective mask in the transform domain, the problem of directional missed detection and grid line misjudgment caused by image blurring in photovoltaic module manufacturing is solved, achieving efficient and accurate defect detection.
Patent Information
- Application Number
- CN202610247674.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-03-02
- Publication Date
- 2026-05-15
AI Technical Summary
In high-speed photovoltaic module manufacturing lines, image motion blur leads to directional omissions and misjudgments of grid structure, affecting the accuracy of deep learning defect detection.
By decomposing a two-dimensional image into a one-dimensional sequence according to the motion direction, performing homomorphic transformation and targeted enhancement, a protective mask is constructed to protect the grid structure features in the transform domain, generating a deblurred enhanced image, and using a deep convolutional neural network for defect identification.
It significantly improves the recall rate for minor defects, avoids damage to the gate structure, reduces computational complexity, solves the problems of directional missed detection and gate breakage misjudgment, and meets the real-time detection requirements of high-speed production lines.
Smart Images
Figure CN122048920A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the technical field of solar cell defect detection, and more particularly to a solar cell defect detection system based on machine vision and deep learning. Background Technology
[0002] In high-speed photovoltaic module manufacturing lines, industrial cameras typically acquire and expose images while the solar cells are in continuous motion to match the production line's pace. This process causes slight linear convolutional blurring in the image along the direction of cell movement, which significantly interferes with deep learning-based automated defect detection systems. Motion blur smooths and weakens high-frequency edge information of subtle defect features perpendicular to the direction of movement, such as microcracks and fine scratches. This results in significant directional missed detection problems when deep learning models identify such defects, i.e., a decrease in sensitivity to defects in specific directions.
[0003] In addition, the surface of solar cells is covered with highly reflective, periodically arranged metal grid lines. Conventional image deblurring or sharpening algorithms are mostly global or indiscriminate enhancements. When attempting to restore image details, they are prone to over-processing the grid line structure, resulting in grid line imaging with breaks, burrs, or deformation. This can lead to false detection of grid breaks in the detection system, seriously affecting the accuracy of detection. Summary of the Invention
[0004] To address the technical problems existing in the background art, this invention proposes a solar cell defect detection system based on machine vision and deep learning, the specific solution of which is as follows: A solar cell defect detection system based on machine vision and deep learning includes: The data acquisition module is used to acquire the original image of the solar cell in motion and its corresponding relative motion direction; The sequence decomposition module is used to convert the original image into multiple independent one-dimensional grayscale sequences based on the relative motion direction. The homomorphic transformation module is used to perform homomorphic transformation on each of the one-dimensional grayscale sequences to generate the corresponding homomorphic transformation feature spectrum. A mask construction module is used to obtain the periodic grid line structure features of a solar cell, and based on the periodic grid line structure features, construct a protective mask in the transform domain corresponding to the homomorphic transform feature spectrum; The inverse transform reconstruction module is used to perform weighted processing on the homomorphic transform feature spectrum based on the protective mask, and to perform inverse homomorphic transformation on the weighted result to generate a reconstructed enhanced one-dimensional grayscale sequence. An image synthesis module is used to generate a deblurred enhanced image based on all the enhanced one-dimensional grayscale sequences; The defect identification module is used to input the enhanced image into a pre-built defect detection model to generate defect detection results for solar cells.
[0005] Furthermore, in the sequence decomposition module, based on the relative motion direction, the original image is converted into multiple independent one-dimensional grayscale sequences, as follows: The direction of relative motion is determined as the direction of image blur expansion; The original image is scanned along a direction perpendicular to the expansion direction to obtain multiple one-dimensional grayscale sequences perpendicular to the expansion direction.
[0006] Furthermore, if the relative motion direction is horizontal, then pixel grayscale values are extracted row by row along the vertical direction to form multiple one-dimensional grayscale sequences in the horizontal direction. If the relative motion direction is vertical, then pixel grayscale values are extracted column by column along the horizontal direction to form multiple one-dimensional grayscale sequences in the vertical direction.
[0007] Furthermore, in the homomorphic transformation module, a homomorphic transformation is performed on each of the one-dimensional gray-level sequences to generate a corresponding homomorphic transformation feature spectrum, as follows: Perform a Fourier transform on the one-dimensional grayscale sequence to generate a frequency domain representation; Calculate the magnitude of the frequency domain representation; Perform a logarithmic operation on the modulus value; Perform an inverse Fourier transform on the result of the logarithmic operation to generate a homomorphic transform feature spectrum.
[0008] Furthermore, an inverse Fourier transform is performed on the result of the logarithmic operation to generate a homomorphic transform feature spectrum, as follows: The homomorphic transform feature spectrum is defined on the pseudo-spatial scaling axis, where: The effects of motion blur manifest as low-frequency components near the pseudo-spatial scale origin; The periodic grid structure is characterized by pulse components that are evenly distributed on the pseudo-spatial scaling axis.
[0009] Furthermore, in the mask construction module, based on the periodic gate line structure features, a protective mask is constructed in the transform domain corresponding to the homomorphic transform feature spectrum, as follows: On the pseudo-spatial scaling axis of the transform domain, the position of the characteristic peak of the corresponding grating structure is identified. The position of the characteristic peak includes the main period characteristic peak of the grating and its higher-order harmonic characteristic peaks. The protected area is defined with the location of the characteristic peak as the center. The enhancement region is defined within the low-frequency region corresponding to motion blur and located outside the protected area; A mask function is constructed such that a first coefficient value is taken within the protected area, a second coefficient value is taken within the enhanced area, and the first coefficient value is taken in the remaining areas, wherein the second coefficient value is less than the first coefficient value.
[0010] Furthermore, the mask function The construction satisfies the following relationship: If q satisfies ,but ; If q satisfies And q does not satisfy ,but ; In other cases ; Where q is the pseudo-spatial scale, P is the scale value corresponding to the physical spacing of the gate lines, k is an integer, and δ is the preset neighborhood radius. The fuzzy kernel influence range scale is α, which is a second coefficient with a value ranging from 0.4 to 0.8.
[0011] Furthermore, in the inverse transform reconstruction module, the homomorphic transform feature spectrum is weighted based on the protective mask, and the weighted result is subjected to an inverse homomorphic transform, as follows: The protective mask is multiplied by the homomorphic transform feature spectrum to obtain a weighted feature spectrum. Perform a Fourier transform on the weighted feature spectrum; Perform exponential operations on the result of the Fourier transform; Perform an inverse Fourier transform on the result of the exponential operation to generate the reconstructed enhanced one-dimensional grayscale sequence.
[0012] Furthermore, in the image synthesis module, based on all the aforementioned enhanced one-dimensional grayscale sequences, a deblurred enhanced image is generated as follows: The enhanced one-dimensional grayscale sequences are arranged and combined according to their corresponding spatial positions in the original image to reconstruct a complete two-dimensional enhanced image, thus generating the enhanced image.
[0013] Furthermore, in the defect identification module, a pre-built defect detection model is constructed: Obtain enhanced image samples of solar cells labeled with various defect types; The enhanced image samples are trained using a deep convolutional neural network to generate the defect detection model; The defect detection model is used to identify the type and location of defects from the input enhanced image, and the defects include at least one of hidden cracks, scratches, and broken grids.
[0014] Compared with the prior art, the present invention can achieve at least the following beneficial effects: 1. This invention decomposes a two-dimensional image into a one-dimensional sequence according to the motion direction, and performs homomorphic transformation and targeted enhancement. This accurately removes linear convolutional blur caused by motion and exposure coupling. By restoring the high-frequency edge information perpendicular to the motion direction that was smoothed out by blurring, directional sensitive defects such as microcracks and fine scratches are clearly revealed again, significantly improving the recall rate for these defects and solving the problem of directional missed detection caused by image blurring. In the homomorphic transformation domain, a protective mask is constructed based on the inherent periodic grid line structure of the solar cell. This mask accurately locates and protects the feature components of the corresponding grid line information in the transformation domain, while deblurring and enhancement are performed in other areas. The system is robust, ensuring that the bright, periodic metal grid structure is fully preserved during image sharpening and defect detail restoration. This avoids grid breakage or deformation caused by conventional global deblurring algorithms, effectively preventing false detections of broken grids and significantly improving detection accuracy. It reduces the two-dimensional deconvolution problem into multiple independent one-dimensional signal processing problems, avoiding the cumbersome and unstable two-dimensional phase calculation and deconvolution process in traditional methods. This not only reduces computational complexity and power consumption but also fundamentally eliminates phase artifacts such as ringing and checkerboard patterns. The entire processing flow is efficient and stable, perfectly meeting the stringent requirements of high-speed photovoltaic production lines for real-time online detection.
[0015] 2. This invention decomposes motion-blurred images into one-dimensional sequences for homomorphic transformation and constructs a protective mask for the grid line period in the transform domain. This achieves the complete protection of the grid line structure while eliminating directional blur and restoring the edges of minor defects. This effectively solves the problems of directional missed detection caused by image blur, false judgment of grid breakage caused by conventional algorithms damaging the grid lines, and the problems of complex calculation and easy artifact generation in traditional two-dimensional restoration methods. It significantly improves the accuracy, robustness and real-time performance of defect detection in high-speed production line environments. Attached Figure Description
[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a system principle block diagram of the present invention. Detailed Implementation
[0017] Embodiments of the present invention are described in detail below. Examples of these embodiments are illustrated in the accompanying drawings, wherein the same or similar symbols denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.
[0018] Reference first Figure 1 This invention provides a solar cell defect detection system based on machine vision and deep learning, comprising: The data acquisition module is used to acquire the original image of the solar cell in motion and its corresponding relative motion direction.
[0019] It should be noted that the original image of the solar cell in motion refers to an image captured by an industrial camera during the continuous movement of the cell on a high-speed photovoltaic production line, without any deblurring processing. Because of the relative motion of the cell during camera exposure, this image will have motion blur.
[0020] It should be noted that the corresponding relative motion direction refers to the direction of movement of the solar cell relative to the industrial camera during exposure. This direction is usually consistent with the running direction of the production line conveyor belt and is known or can be directly obtained from parameters of the production line control system (such as encoders, PLCs, etc.).
[0021] The sequence decomposition module is used to convert the original image into multiple independent one-dimensional grayscale sequences based on the relative motion direction.
[0022] In an optional embodiment, the sequence decomposition module converts the original image into multiple independent one-dimensional grayscale sequences based on the relative motion direction, as follows: The direction of relative motion is determined as the direction of image blur expansion; It should be noted that determining the relative motion direction as the expansion direction of the image blur is based on the physical cause of motion blur: the blur of each pixel in the image is the result of integration along the motion path, therefore the direction of blur diffusion (convolution) is the relative motion direction. Determining this expansion direction is to clarify the dimension of the image dimensionality reduction process.
[0023] The original image is scanned along a direction perpendicular to the expansion direction to obtain multiple one-dimensional grayscale sequences perpendicular to the expansion direction.
[0024] It should be noted that motion blur only causes convolution effects in the expansion direction, while in the direction perpendicular to the expansion direction (i.e., the scanning direction), image information is independent and not blurred. By extracting pixel values row by row (or column by column) along this vertical direction, several one-dimensional grayscale sequences are obtained. Each sequence is essentially a grayscale distribution along a line on the surface of the solar cell along the scanning direction, and the order of the sequences corresponds to their spatial positions in the expansion direction.
[0025] In an optional embodiment, if the relative motion direction is horizontal, then pixel grayscale values are extracted row by row along the vertical direction to form multiple horizontal one-dimensional grayscale sequences. If the relative motion direction is vertical, then pixel grayscale values are extracted column by column along the horizontal direction to form multiple one-dimensional grayscale sequences in the vertical direction.
[0026] It should be noted that the complex problem of two-dimensional image deconvolution is simplified into a series of independent one-dimensional signal enhancement problems. This avoids the unstable two-dimensional phase deconvolution process that must be handled in traditional two-dimensional deblurring algorithms, thus eliminating the resulting ringing artifacts and checkerboard noise, while greatly reducing computational complexity.
[0027] It should be noted that in practice, if the relative motion is known to be horizontal (i.e., the battery cells are moving horizontally), then the system scans row by row along the vertical direction to obtain a series of one-dimensional grayscale sequences in the horizontal direction (i.e., the row direction). Conversely, if the relative motion is vertical, then the system scans column by column along the horizontal direction to obtain a series of one-dimensional grayscale sequences in the vertical direction. Here, "horizontal" and "vertical" refer to the image coordinate system; in practical applications, the production line layout and camera installation orientation are taken into account.
[0028] The homomorphic transformation module is used to perform homomorphic transformation on each of the one-dimensional grayscale sequences to generate the corresponding homomorphic transformation feature spectrum.
[0029] It should be noted that the homomorphic transform is a special signal processing technique that transforms mixed signals, originally coupled together through multiplication or convolution, into a form that can be added and separated in the transform domain. For solar cell images, the imaging process can be approximated as: the clear cell texture is convolved with the camera's point spread function (represented in this case as a motion blur kernel). The homomorphic transform converts the convolution relationship into an additive relationship through logarithmic operations, thus enabling the subsequent separation of blurred components from the true texture components.
[0030] In an optional embodiment, the homomorphic transformation module performs a homomorphic transformation on each of the one-dimensional gray-level sequences to generate a corresponding homomorphic transformation feature spectrum, as follows: Perform a Fourier transform on the one-dimensional grayscale sequence to generate a frequency domain representation; Calculate the magnitude of the frequency domain representation; Perform a logarithmic operation on the modulus value; Perform an inverse Fourier transform on the result of the logarithmic operation to generate a homomorphic transform feature spectrum.
[0031] It should be noted that the sequence of operations performed on each one-dimensional grayscale sequence follows a path from Fourier transform, modulo operation, logarithmic operation to inverse Fourier transform. Specifically: Fourier transform: converts a spatial domain signal to the frequency domain. Motion blur in the frequency domain manifests as low-frequency attenuation, meaning that blur weakens the high-frequency components of an image (such as edges).
[0032] Modulus extraction: Obtain the amplitude spectrum of the frequency domain signal and discard phase information. In this method, we mainly utilize amplitude information to analyze and separate signal components.
[0033] Logarithmic operation: This is the core step of the homomorphic transformation. It transforms the multiplicative modulation relationship caused by convolution (blurring) in the frequency domain amplitude spectrum (approximately represented as the product of the fuzzy kernel spectrum and the texture spectrum) into a simple additive relationship. That is: log(fuzzy texture spectrum) ≈ log(fuzzy kernel spectrum) + log(texture spectrum).
[0034] Inverse Fourier Transform: This transforms the logarithmic amplitude spectrum back to a new domain. This domain is not the original spatial domain, but an analytical domain with special physical meaning generated after logarithmic operations and a second transformation. This invention defines it as a pseudo-spatial scaling axis. In this domain, the different physical source components of the signal exhibit distinguishable distribution characteristics.
[0035] In an optional embodiment, an inverse Fourier transform is performed on the result of the logarithmic operation to generate a homomorphic transform feature spectrum, as follows: The homomorphic transform feature spectrum is defined on the pseudo-spatial scaling axis, where: The effects of motion blur manifest as low-frequency components near the pseudo-spatial scale origin; The periodic grid structure is characterized by pulse components that are evenly distributed on the pseudo-spatial scaling axis.
[0036] It should be noted that the pseudo-spatial scaling axis is obtained by performing an inverse Fourier transform on the logarithmic amplitude spectrum. Its dimensions differ from the original spatial domain, but it retains the excellent property of component separability. On this axis: The effects of motion blur manifest as low-frequency, slowly varying components concentrated near the pseudo-spatial scale origin. This is because the motion blur kernel typically has smooth, low-pass properties, and after homomorphic transformation, its main energy is concentrated in the low-scaling region of the pseudo-spatial domain.
[0037] The periodic grid line structure is characterized by sharp pulse components that are evenly spaced along the pseudo-spatial scaling axis. This is because the metal grid lines on the surface of the solar cell have strict spatial periodicity. After passing through the aforementioned homomorphic transformation chain, this periodic signal will produce a significant impulse response at the scaling position corresponding to its period (and its integer multiple harmonic positions), forming a series of easily identifiable characteristic peaks.
[0038] It is worth noting that defining the homomorphic transform feature spectrum on the pseudo-spatial scaling axis and clarifying its component characteristics is of significant technical importance. It allows us to directly and clearly distinguish harmful blurring components from the periodic components of the grating lines that need protection, based on differences in position and shape, within the transform domain. This provides a direct and reliable basis for the subsequent accurate construction of the protective mask, enabling the enhancement algorithm to precisely suppress blur (enhancement region) while completely preserving the grating line structure (protection region). This fundamentally solves the industry problem of conventional deblurring or sharpening algorithms easily damaging the grating lines and causing false detections.
[0039] A mask construction module is used to obtain the periodic grid line structure features of a solar cell and, based on the periodic grid line structure features, construct a protective mask in the transform domain corresponding to the homomorphic transform feature spectrum.
[0040] It should be noted that obtaining the periodic grid line structure characteristics of solar cells does not refer to reading a fixed parameter from an external database, but rather to automatically analyzing, identifying, and extracting a set of key information that characterizes the periodicity of the grid lines from the homomorphic transformation feature spectrum obtained in the previous step. This process is based on morphological analysis of the feature spectrum on the pseudo-spatial scaling axis.
[0041] In an optional embodiment, the mask construction module constructs a protective mask in the transform domain corresponding to the homomorphic transform feature spectrum based on the periodic gate line structure features, as follows: On the pseudo-spatial scaling axis of the transform domain, the position of the characteristic peak of the corresponding grating structure is identified. The position of the characteristic peak includes the main period characteristic peak of the grating and its higher-order harmonic characteristic peaks. It should be noted that identifying the characteristic peak positions of the corresponding grating structure on the pseudo-spatial scaling axis of the transform domain specifically refers to searching for sharp pulse signals with amplitudes significantly higher than the background noise on the pseudo-spatial scaling axis of the homomorphic transform characteristic spectrum. Due to the strict periodicity of the grating lines, these pulse signals will appear at a series of equally spaced positions. Among them, the pulse with the highest amplitude corresponding to the fundamental physical period of the grating line is identified as the main periodic characteristic peak of the grating line (corresponding to k=1). Pulses appearing at integer multiples of its scaling position (k=2,3,…) with amplitudes that typically decrease sequentially are identified as higher-order harmonic characteristic peaks. The set of these characteristic peak positions completely defines the periodicity of the grating line.
[0042] The protected area is defined with the location of the characteristic peak as the center. It should be noted that defining the protection zone centered on the characteristic peak position is to ensure that all key frequency components (including the dominant frequency and higher harmonics) constituting the grating structure are fully preserved in subsequent processing, avoiding thinning, breakage, or burr formation in the grating lines. The protection zone is typically set as a small neighborhood (±δ) before and after each characteristic peak position. The radius δ of this neighborhood needs to be reasonably set according to the signal sampling rate and spectral resolution to ensure coverage of the energy core of the characteristic peak.
[0043] The enhancement region is defined within the low-frequency region corresponding to motion blur and located outside the protected area; It should be noted that determining the enhancement region range, within the low-frequency region corresponding to motion blur and located outside the protected area, is one of the core innovations of this invention. The technical principle is that, after homomorphic transformation, the energy of motion blur is mainly concentrated in the low-frequency region near the pseudo-spatial scaling origin (the range where the scaling value q is relatively small, for example...). The enhancement region is set to precisely locate the area where the blur components that need to be suppressed are located. At the same time, the protection zone is excluded from the enhancement region to ensure that the enhancement operation (i.e., attenuating the blur components) does not affect the raster information, thus achieving precise processing.
[0044] A mask function is constructed such that a first coefficient value is taken within the protected area, a second coefficient value is taken within the enhanced area, and the first coefficient value is taken in the remaining areas, wherein the second coefficient value is less than the first coefficient value.
[0045] In an optional embodiment, the mask function The construction satisfies the following relationship: If q satisfies ,but ; If q satisfies And q does not satisfy ,but ; In other cases ; Where q is the pseudo-spatial scale, P is the scale value corresponding to the physical spacing of the gate lines, k is an integer, and δ is the preset neighborhood radius. The fuzzy kernel influence range scale is α, which is a second coefficient with a value ranging from 0.4 to 0.8.
[0046] It should be noted that constructing a mask function is a mathematical tool for achieving selective enhancement. This function acts directly on the homomorphic transform eigenspectrum, and its coefficients determine whether the corresponding scaling components are preserved, attenuated, or unchanged.
[0047] First coefficient value (usually set to 1.0): Used in the protected area and the rest of the area. A coefficient of 1.0 means full pass, that is, the energy of the corresponding scale component (raster information and most normal texture) is 100% preserved, achieving zero-damage protection of the raster structure.
[0048] The second coefficient value is the value of the second coefficient α (range 0.4-0.8): used in the enhancement region. A coefficient less than 1 means that the signal components located in this region are attenuated to a certain extent. Since this region mainly corresponds to blurred components, attenuating it is equivalent to inversely compensating for the high-frequency loss caused by blurring in the transform domain, thereby achieving edge sharpening and deblurring effects in the image spatial domain.
[0049] It should be noted that for the range of values for α, the lower limit of effectiveness (α < 0.4) is: if the value of α is too low, the suppression of the blur component is excessive. After the inverse transformation, obvious ringing artifacts and amplified background noise will appear in the image. These artifacts are very likely to be misjudged as real defects by subsequent defect detection models, resulting in a significant increase in the false alarm rate.
[0050] Upper limit of effect (α>0.8): If the α value is too high, the suppression of blur is insufficient and it cannot effectively restore the edge contrast of the small defects that have been blurred (especially the hidden cracks perpendicular to the direction of motion), resulting in poor deblurring effect and failing to achieve the fundamental goal of improving defect recall rate.
[0051] It should be noted that the mask function The mathematical definition of is a precise expression of the protection and enhancement logic described above. The parameter P (the scale value corresponding to the gate period) and The influence range of the blur kernel can be set based on prior knowledge or automatically estimated by analyzing the feature spectrum. The integer k covers all harmonic orders that need to be protected. This function is concise yet powerful, transforming complex image processing problems into a single point-by-point multiplication operation in the transform domain, resulting in extremely high computational efficiency that fully meets the real-time requirements of industrial production lines. Through the weighted processing of this mask, this invention successfully unifies the two contradictory goals of blur suppression and grid protection in traditional methods.
[0052] The inverse transform reconstruction module is used to perform weighted processing on the homomorphic transform feature spectrum based on the protective mask, and to perform inverse homomorphic transformation on the weighted result to generate a reconstructed enhanced one-dimensional grayscale sequence.
[0053] It should be noted that the image selective enhancement is achieved by weighting the homomorphic transform feature spectrum based on the protective mask. Here, the weighting specifically refers to applying weights to the mask function constructed by the mask construction module. The signal is multiplied point-by-point by the one-dimensional homomorphic transform feature spectrum generated by the homomorphic transform module. The mathematical essence of this operation is amplitude modulation of different components of the signal in the transform domain. In the area with a mask factor of 1.0 (protected areas and other areas), the signal components of the corresponding scale are preserved intact, ensuring that the grid line information and most normal textures are not affected; In the region where the masking factor is α (less than 1) (enhancement region), the signal components corresponding to the scale are attenuated proportionally. Since this region has been identified as the core area affected by motion blur, attenuating these components is equivalent to weakening the intensity of the blurred signal in the transform domain; Through this precise weighting, the fuzzy components are effectively suppressed while the useful signals are preserved, thus initially completing the tasks of defuzzification and structure protection in the transform domain.
[0054] In an optional embodiment, in the inverse transform reconstruction module, the homomorphic transform feature spectrum is weighted based on the protective mask, and the weighted result is subjected to an inverse homomorphic transform, as follows: The protective mask is multiplied by the homomorphic transform feature spectrum to obtain a weighted feature spectrum. Perform a Fourier transform on the weighted feature spectrum; Perform exponential operations on the result of the Fourier transform; Perform an inverse Fourier transform on the result of the exponential operation to generate the reconstructed enhanced one-dimensional grayscale sequence.
[0055] It should be noted that the inverse homomorphic transformation of the weighted result is a strict inverse sequence of operations corresponding to the forward homomorphic transformation of the homomorphic transformation module. The purpose is to completely and correctly reconstruct the processed transform domain signal back into a one-dimensional grayscale sequence in the spatial domain. This sequence specifically includes: Perform a Fourier transform on the weighted characteristic spectrum: transform the modulated signal, which is located on the pseudo-spatial scaling axis, back into the frequency domain. This is to restore the logarithmic amplitude representation of the signal in the frequency domain, preparing for subsequent exponential inverse operations.
[0056] The Fourier transform result is then subjected to an exponential operation: this performs the exact opposite mathematical operation to the logarithmic operation in the forward transform. While the logarithmic operation reduces the convolution relationship to addition, the exponential operation transforms the signal, which has already undergone weighted subtraction of ambiguity components in the transform domain, back into a multiplicative relationship in the frequency domain. After this step, we obtain the corrected (with ambiguity components suppressed) frequency domain amplitude spectrum.
[0057] An inverse Fourier transform is performed on the result of the exponential operation: the corrected frequency domain signal is transformed back to the original spatial domain, thereby generating a reconstructed enhanced one-dimensional grayscale sequence. Compared to the original, blurred one-dimensional grayscale sequence, this sequence differs in that the edge diffusion and contrast reduction caused by motion blur are significantly recovered, especially the high-frequency edge information of tiny defects (such as microcracks) perpendicular to the direction of motion are re-highlighted; while the grid structure parallel to the direction of motion, because its periodic components are fully protected in the transform domain, is perfectly preserved in this sequence without damage.
[0058] It should be noted that the inverse transform reconstruction module achieves a closed loop from analysis to synthesis. It receives a protective mask representing the problem analysis results and a homomorphic transform feature spectrum, and outputs a corrected signal through a precise inverse transform process. All complex processing is completed within the one-dimensional transform domain of the signal via linear or dot product operations, completely avoiding the unavoidable and unstable phase estimation and iterative optimization processes in traditional two-dimensional image restoration algorithms. This fundamentally eliminates ringing artifacts and achieves extremely high computational efficiency, laying a high-quality image foundation for subsequent real-time deep learning defect detection.
[0059] The image synthesis module is used to generate a deblurred enhanced image based on all the enhanced one-dimensional grayscale sequences.
[0060] In an optional embodiment, the image synthesis module generates a deblurred enhanced image based on all the enhanced one-dimensional grayscale sequences, as follows: The enhanced one-dimensional grayscale sequences are arranged and combined according to their corresponding spatial positions in the original image to reconstruct a complete two-dimensional enhanced image, thus generating the enhanced image.
[0061] It should be noted that generating the deblurred enhanced image based on all the aforementioned enhanced one-dimensional grayscale sequences is an image reconstruction and synthesis process. Each enhanced one-dimensional grayscale sequence corresponds to a line (row or column) in the original two-dimensional image that is perpendicular to the motion blur direction, and has already independently completed motion blur correction and protective enhancement for the grid structure in the preceding steps.
[0062] It should be noted that arranging and combining these processed, discrete one-dimensional signals according to their corresponding spatial positions in the original image means repositioning them to their corresponding spatial positions based on their row or column order in the original image before decomposition by the sequence decomposition module. This process is essentially the reverse of the decomposition operation of the sequence decomposition module. Specifically, if the initial decomposition was by row, each processed sequence is now treated as a row and stacked sequentially to reconstruct the height dimension of the image; if the initial decomposition was by column, they are arranged side-by-side sequentially to reconstruct the width dimension of the image.
[0063] It should be noted that the two-dimensional enhanced image generated through this decomposition-independent processing-reconstruction process has the following decisive improvements compared to the original blurred image: Directional blurring is effectively suppressed: Because each one-dimensional sequence is independently deblurred and enhanced, the edge contrast and detail sharpness of the entire image in the direction of motion are systematically restored. This makes minute defects perpendicular to the direction of motion (such as microcracks) that were originally difficult to detect due to blurring clearly visible.
[0064] The integrity of the grid line structure is guaranteed: Since a protective mask is used when processing each sequence independently, all grid lines in the reconstructed image retain their original shape, width and continuity, effectively avoiding grid line breakage or deformation that may be caused by conventional global filtering.
[0065] Artifact-free stitching: Since the processing method is based on one-dimensional signal processing, it completely avoids the complex phase alignment problem in two-dimensional deconvolution. Moreover, the processing of each sequence is independent of each other. Therefore, the reconstructed image will not introduce artifacts such as stitching marks, ringing, or checkerboard patterns at the sequence boundaries, and the image quality is uniform and natural.
[0066] The defect identification module is used to input the enhanced image into a pre-built defect detection model to generate defect detection results for solar cells.
[0067] In an optional embodiment, a defect detection model is pre-built in the defect identification module: Obtain enhanced image samples of solar cells labeled with various defect types; The enhanced image samples are trained using a deep convolutional neural network to generate the defect detection model; The defect detection model is used to identify the type and location of defects from the input enhanced image, and the defects include at least one of hidden cracks, scratches, and broken grids.
[0068] It should be noted that the pre-built defect detection model is a model with stable defect recognition capabilities, pre-trained offline using a large amount of labeled data. The labeled enhanced image samples of solar cells with various defect types form the basis of the model training. These samples are enhanced images obtained by processing the original blurred images. Each enhanced image is precisely labeled by professionals, indicating the specific category of the defects (such as hidden cracks, scratches, and broken grids) and their precise location in the image (usually represented by bounding boxes or polygons). Using enhanced images instead of the original blurred images for training is key to achieving high accuracy, as it provides the model with clear, unblurred, and complete positive samples, enabling the model to learn the most essential visual features of the defects.
[0069] It should be noted that training the enhanced image samples using a deep convolutional neural network refers to leveraging the powerful feature extraction, classification, and localization capabilities of convolutional neural networks. During training, the network model learns from a large number of samples how to automatically identify texture, edge, shape, and contextual feature patterns related to defect categories such as hidden cracks, scratches, and broken grids from the pixel data of the enhanced images. The trained model encapsulates this learned knowledge.
[0070] It should be noted that inputting the enhanced image generated by the image synthesis module for the current battery cell into this pre-trained model is the final step in the pipeline. The model performs forward inference on the input image, and its output is the defect detection result, which typically includes: Defect Category: Identify the types of defects present in the image.
[0071] Defect location: The specific area of each defect in the image is given in coordinate form.
[0072] Confidence level: The degree to which the model is confident about each recognition result.
[0073] It should be noted that this invention combines image enhancement based on a physical model with data-driven deep learning detection, forming a complete technical solution with complementary advantages. Front-end image enhancement solves the problems of blurring and damage to the grid lines, providing high-quality input for the back-end deep learning model; the back-end deep learning model provides efficient, accurate, and scalable defect recognition capabilities. This combination enables the entire system to achieve real-time detection of defects as small as micrometers, such as microcracks, with high recall and low false alarm rates in high-speed production line environments, meeting the dual needs of the photovoltaic manufacturing industry for quality control and capacity improvement.
[0074] In summary, this patent provides a defect detection method that can efficiently eliminate directional motion blur, accurately protect critical grid structures, and ensure processing speed and image quality in a high-speed production line environment, thus achieving a comprehensive improvement in detection accuracy and reliability.
Claims
1. A solar cell defect detection system based on machine vision and deep learning, characterized in that, include: The data acquisition module is used to acquire the original image of the solar cell in motion and its corresponding relative motion direction; The sequence decomposition module is used to convert the original image into multiple independent one-dimensional grayscale sequences based on the relative motion direction. The homomorphic transformation module is used to perform homomorphic transformation on each of the one-dimensional grayscale sequences to generate the corresponding homomorphic transformation feature spectrum. A mask construction module is used to obtain the periodic grid line structure features of a solar cell, and based on the periodic grid line structure features, construct a protective mask in the transform domain corresponding to the homomorphic transform feature spectrum; The inverse transform reconstruction module is used to perform weighted processing on the homomorphic transform feature spectrum based on the protective mask, and to perform inverse homomorphic transformation on the weighted result to generate a reconstructed enhanced one-dimensional grayscale sequence. An image synthesis module is used to generate a deblurred enhanced image based on all the enhanced one-dimensional grayscale sequences; The defect identification module is used to input the enhanced image into a pre-built defect detection model to generate defect detection results for solar cells.
2. The solar cell defect detection system based on machine vision and deep learning as described in claim 1, characterized in that: In the sequence decomposition module, based on the relative motion direction, the original image is converted into multiple independent one-dimensional grayscale sequences, as follows: The direction of relative motion is determined as the direction of image blur expansion; The original image is scanned along a direction perpendicular to the expansion direction to obtain multiple one-dimensional grayscale sequences perpendicular to the expansion direction.
3. The solar cell defect detection system based on machine vision and deep learning as described in claim 2, characterized in that: If the relative motion direction is horizontal, then the pixel grayscale values are extracted row by row along the vertical direction to form multiple one-dimensional grayscale sequences in the horizontal direction. If the relative motion direction is vertical, then pixel grayscale values are extracted column by column along the horizontal direction to form multiple one-dimensional grayscale sequences in the vertical direction.
4. The solar cell defect detection system based on machine vision and deep learning as described in claim 1, characterized in that: In the homomorphic transformation module, a homomorphic transformation is performed on each of the one-dimensional gray-level sequences to generate the corresponding homomorphic transformation feature spectrum, as follows: Perform a Fourier transform on the one-dimensional grayscale sequence to generate a frequency domain representation; Calculate the magnitude of the frequency domain representation; Perform a logarithmic operation on the modulus value; Perform an inverse Fourier transform on the result of the logarithmic operation to generate a homomorphic transform feature spectrum.
5. The solar cell defect detection system based on machine vision and deep learning as described in claim 4, characterized in that: Performing an inverse Fourier transform on the result of the logarithmic operation generates a homomorphic transform characteristic spectrum, as follows: The homomorphic transform feature spectrum is defined on the pseudo-spatial scaling axis, where: The effects of motion blur manifest as low-frequency components near the pseudo-spatial scale origin; The periodic grid structure is characterized by pulse components that are evenly distributed on the pseudo-spatial scaling axis.
6. The solar cell defect detection system based on machine vision and deep learning as described in claim 1, characterized in that: In the mask construction module, based on the periodic gate line structure features, a protective mask is constructed in the transform domain corresponding to the homomorphic transform feature spectrum, as follows: On the pseudo-spatial scaling axis of the transform domain, the position of the characteristic peak of the corresponding grating structure is identified. The position of the characteristic peak includes the main period characteristic peak of the grating and its higher-order harmonic characteristic peaks. The protected area is defined with the location of the characteristic peak as the center. The enhancement region is defined within the low-frequency region corresponding to motion blur and located outside the protected area; A mask function is constructed such that a first coefficient value is taken within the protected area, a second coefficient value is taken within the enhanced area, and the first coefficient value is taken in the remaining areas, wherein the second coefficient value is less than the first coefficient value.
7. The solar cell defect detection system based on machine vision and deep learning as described in claim 6, characterized in that: The mask function The construction satisfies the following relationship: If q satisfies ,but ; If q satisfies And q does not satisfy ,but ; In other cases ; Where q is the pseudo-spatial scale, P is the scale value corresponding to the physical spacing of the gate lines, k is an integer, and δ is the preset neighborhood radius. The fuzzy kernel influence range scale is α, which is a second coefficient with a value ranging from 0.4 to 0.
8.
8. The solar cell defect detection system based on machine vision and deep learning as described in claim 1, characterized in that: In the inverse transform reconstruction module, the homomorphic transform feature spectrum is weighted based on the protective mask, and the weighted result is then subjected to an inverse homomorphic transform, as follows: The protective mask is multiplied by the homomorphic transform feature spectrum to obtain a weighted feature spectrum. Perform a Fourier transform on the weighted feature spectrum; Perform exponential operations on the result of the Fourier transform; Perform an inverse Fourier transform on the result of the exponential operation to generate the reconstructed enhanced one-dimensional grayscale sequence.
9. The solar cell defect detection system based on machine vision and deep learning as described in claim 1, characterized in that: In the image synthesis module, based on all the enhanced one-dimensional grayscale sequences, a deblurred enhanced image is generated, as follows: The enhanced one-dimensional grayscale sequences are arranged and combined according to their corresponding spatial positions in the original image to reconstruct a complete two-dimensional enhanced image, thus generating the enhanced image.
10. The solar cell defect detection system based on machine vision and deep learning as described in claim 1, characterized in that: In the defect identification module, a pre-built defect detection model is used: Obtain enhanced image samples of solar cells labeled with various defect types; The enhanced image samples are trained using a deep convolutional neural network to generate the defect detection model; The defect detection model is used to identify the type and location of defects from the input enhanced image, and the defects include at least one of hidden cracks, scratches, and broken grids.