Optics and Precision Engineering, Volume. 32, Issue 6, 868(2024)
Review of defect detection algorithms for solar cells based on machine vision
Figures & Tables(22)
Fig. 1. Diagram of the entire process of crystalline silicon solar cell production
Fig. 2. Surface images of solar cells obtained with different imaging methods
Fig. 3. Framework of surface defect detection algorithm of solar cells based on traditional machine vision
Fig. 4. Framework of traditional machine vision solar cell defect detection algorithms
Fig. 5. Textural differences in edge gradients between cracks， broken grid defects and the background
Fig. 6. Diagram of the journey of development of deep learning networks
Table 1. Common defects of solar cells
View tableView in ArticleTable 1. Common defects of solar cells
Defect category Defect name Impact on cell performance Visual
characteristics
Causes Sampling Shape defects Cracks Resulting in the interruption of the current path， reducing the conversion efficiency of the cell Tiny hidden cracks or larger cracks Welding or cutting errors， production process errors， impacts Broken gate Broken grid part can not collect the current， resulting in lower cell conversion efficiency Interrupted or broken conductive grid lines Missing corners，Broken May cause damage to the grid line， affecting the module power factor Missing or redundant shape Color defects Spots、Black spots May reduce the absorption of light by the solar cell， lowering the photovoltaic conversion efficiency Abnormal or uneven surface coloration Uneven chemical reaction during coating， defective raw silicon material Black flakes Failure of the black part of the solar cell to provide power may cause thermal breakdown， resulting in lower output power of the whole module Texture defects Water marks，Fingerprints，
Wheel marks
Resulting in uneven reflection or absorption of light from the surface and a decrease in photoelectric conversion efficiency in some areas Areas of abnormal brightness， speckles or fingerprints， wheel marks on the surface Improper manual operation， excessive machine pressure， untidy production environment
Table 2. Advantages and disadvantages of specific algorithms based on traditional machine vision
View tableView in ArticleTable 2. Advantages and disadvantages of specific algorithms based on traditional machine vision
Method Advantages Limitations Improved anisotropic diffusion filter［ 24 ］Accurate segmentation of hidden cracks Only applicable for microcrack detection Mean shift technique［ 25 ］Significant detection of fingerprint and contamination defects Method lacks generality RBF+ SVM［ 34 ］Detection of multiple defects， strong
interference resistance
Does not meet real-time production inspection
requirements
KAZE+SVM［ 35 ］Low resource consumption， relatively short processing time Lower accuracy on polycrystalline units DBF+SVM［ 41 ］Good performance on imbalanced
datasets， fast detection speed
Limited labeled image data in the dataset Uniform distribution clustering［ 45 ］Improved detection accuracy for various types of crack defects Limited to small datasets， has limitations on the size of the inspection images CPICS-LBP+SVM［ 47 ］Extraction of defect features in the
presence of uneven background
Only considers the average of single-layer adjacent pixel information Visual attention mechanism［ 51 ］Accurate localization of defect areas by the model Does not meet real-time monitoring production
requirements
PSO+SVM［ 52 ］Strong interference resistance，
high generality
Limited samples in the dataset ICA［ 58 ］Relatively high detection accuracy Unable to distinguish between grid breakage and
microcracks
Table 3. Comparison of various algorithms based on traditional machine vision
View tableView in ArticleTable 3. Comparison of various algorithms based on traditional machine vision
Algorithm Advantages Disadvantages Applicable image types Image domain analysis Gradient features Relatively simple algorithm， suitable for cases where defects and background have significant brightness differences Complex detection process， limited adaptability to illumination changes EL images Other features Comprehensive consideration of multiple features， improving detection robustness Complex feature selection and combination process， requires sufficient domain knowledge Clustering method Automatically discovers clustering structures in data Clustering parameters need targeted real-time adjustments Thresholding method Simple implementation， suitable for simple scenes Clustering parameters need targeted real-time adjustments Region growing method Obtains local structures， suitable for continuous defect areas Sensitive to the selection of initial seeds， challenging for complex-shaped defects Transform domain analysis Matrix decomposition method Better differentiation of different types of defects Higher requirements for image acquisition Digital images Fourier transform Good performance for scratches， cracks， spots， and grid breakage， low complexity， good real-time performance Higher requirements for image preprocessing， parameter setting needed Wavelet transform ICA Good performance for scratches， cracks， spots， and grid breakage， low complexity， good real-time performance Algorithm performance heavily relies on data preprocessing
Table 4. Advantages and disadvantages of specific methods of supervised learning
View tableView in ArticleTable 4. Advantages and disadvantages of specific methods of supervised learning
Network Method Year Advantages Limitations Detection ResNet50+Faster R-CNN+GA-RPN［ 85 ］2019 High detection accuracy， fast speed Incomplete coverage of battery types and defect types in the dataset Faster RPAN-CNN［ 87 ］2020 Good generality in detecting multiple defects Suboptimal detection for multi-scale defects MF-RPN+Faster R-CNN［ 84 ］2021 Strong adaptability to defect shape and scale changes Long detection time， high algorithm complexity BAFPN-CNN［ 88 ］2021 Robust detection of solar cell crack scales Manual setting of feature balance factor in attention module DPiT+CW-MSA［ 95 ］2021 Strong feature extraction capability， good detection performance Relatively large computational load BPGA-CNN［ 91 ］2022 Reduces the impact of complex background noise on small target defect detection Limited types of defects in the dataset YOLO v5+CSP+ECA-Net［ 92 ］2022 High detection accuracy， good performance High model complexity， slow detection speed Densenet121+YOLOv4［ 90 ］2023 Improved detection accuracy and speed Model performance needs further improvement YOLO v5+DCNv2+CA［ 93 ］2023 Good detection performance for small-sized defects， fewer model parameters Uneven distribution of defect types in the dataset DCNN［ 94 ］2023 Quantifies the degree of defects， making the solar cell manufacturing process more intelligent Low resolution of images used due to poor image acquisition system quality Segmentation MAU-Net［ 99 ］2020 Effectively extracts salient features， good segmentation performance Requires a well-annotated training dataset Pre-trained U-net［ 101 ］2020 Directly obtains defect segmentation maps Decreased pixel-level accuracy， longer processing time ERDCF-Net［ 97 ］2022 Effectively suppresses heterogenous texture background interference， grades solar cells based on damage level Uses only crack defect size as a reference for battery quality grading U2-Net+CSEB+DE-RSU［ 100 ］2023 Finer segmentation of defect edges Lower resolution of images obtained using PL imaging technology Hybrid VGG16+U-Net++［ 103 ］2021 Avoids overfitting， combines detection and segmentation tasks Improvement needed in defect localization effectiveness Faster R-CNN+EfficientNet+AE［ 104 ］2021 Implements complete end-to-end network Model robustness needs improvement Lightweight ShuffleNet V2［ 105 ］2022 Improved detection performance， practical utility Poor detection performance for individual defect types YOLOv3-Tiny ［107］ 2022 Balances detection accuracy and speed， achieves real-time monitoring Weak model generalization ability Inception V3 ［106］ 2023 High accuracy in defect recognition， fast classification speed Low pixel count in dataset images NAS+KD［ 108 ］2023 Low hardware requirements， high detection accuracy， meets industrial application need Classifies only whether a battery has defects， cannot determine defect location and type
Table 4. Advantages and disadvantages ofw specific methods of supervised learning
View tableView in ArticleTable 4. Advantages and disadvantages ofw specific methods of supervised learning
Network Method Year Advantages Limitations Classification VGG-16+SVM［ 83 ］2018 Suitable for small datasets Poor real-time detection Transfer Learning -VGG-19-［ 35 ］2019 Good accuracy on single crystal and uneven polycrystalline units Only identifies defects， does not distinguish defect types Improved LeNet-5［ 62 ］2019 Higher accuracy and lower cross-entropy loss than classical LeNet-5 and SVM classifier Limited samples， focuses on one type of defect （cracks） SEF-CNN［ 63 ］2019 Features with better discriminability and robustness Longer model training time ADI［ 71 ］2020 Detects defect types while judging the defect samples Relatively low accuracy Modified AlexNet［ 73 ］2020 High accuracy for large-area defect detection， strong generality Only applicable to visible defects， not suitable for minor scratch detection HFCNN［ 75 ］2020 Low parameter usage， low memory occupation， eliminates data uncertainty Relatively slow training speed L-CNN［ 41 ］2021 Simple network structure， fast training speed Imbalanced class distribution in the dataset， model accuracy needs improvement Ensemble Learning CNN［ 68 ］2021 Improves detection accuracy while main-taining low cost Detection accuracy and speed are not high enough PSO Pruner-CNN［ 80 ］2022 Reduces model complexity Slight decrease in model accuracy
Table 5. Advantages and disadvantages of unsupervised learning， weakly supervised and semi-supervised learning
View tableView in ArticleTable 5. Advantages and disadvantages of unsupervised learning， weakly supervised and semi-supervised learning
Algorithm Year Advantages Limitations Unsupervised learning 2014［ 17 ］Fast detection speed， some level of generality Small dataset， low image resolution 2015［ 112 ］Avoids the impact of manually designed model parameters， strong interference resistance， high generality Limited samples in the dataset 2018［ 113 ］Exhibits a certain level of generality Long training time， unable to handle high-resolution images 2020［ 114 ］Model exhibits some robustness， strong feature expression capability Weak model generalization capability 2020［ 115 ］Demonstrates good generality Cannot directly detect defects in the original image Weakly super-vised learning 2019［ 116 ］Effectively segments based on a small number of annotations Limited annotation information， rough segmentation results 2020［ 117 ］Robustness and adaptability to complex surfaces without the need for extensive annotation data， has some generality Limited annotation information， rough segmentation results 2022［ 118 ］Model training requires only a small number of negative samples， strong generalization capability Does not classify defect types
Table 6. Characteristics of various algorithms based on deep learning
View tableView in ArticleTable 6. Characteristics of various algorithms based on deep learning
Method Network Output result Network characteristics Evaluation Supervised Classification Defect category Only suitable for detecting the presence of single or multiple types of defects in solar cells， relatively low annotation cost The algorithm has strong learning ability for high-dimensional features， but computational complexity increases with the depth of the network Detection Defect location Suitable for classifying and locating multiple defects in a single image， higher annotation cost than classification networks Segmentation Defect shape Suitable for classifying and locating multiple types of defects， pixel-level classification， provides specific defect shapes， higher annotation cost than classification and detection networks Unsupervised GAN Defect existence Globally trains the entire network， usability needs improvement Saves manual annotation costs， can detect unknown defects， but cannot classify defect types DBN Defect feature weights Can generate data， network pre-trained layer by layer， strong feature extraction ability， but network parameters are limited by experience AE Predicted Mask Strong defect feature representation， higher robustness and accuracy than supervised learning， requires consistent input-output data dimensions Weakly supervised and semi-supervised － Defect category or location Localization performance is worse than supervised methods but better than unsupervised methods Suitable for situations with imprecise data
Table 7. Solar cell surface defect detection public dataset
View tableView in ArticleTable 7. Solar cell surface defect detection public dataset
Dataset Size Link 光伏异常检测 392 https：//aistudio.baidu.com/aistudio/datasetdetail/183793 PTID 3 027 http：//vrai.dii.univpm.it/content/photovoltaic-thermal-images-dataset ELPV 2 624 https：//github.com/zae-bayern/elpv-dataset ELDDS1400c5 1 400 https：//universe.roboflow.com/fma04-fayoum-edu-eg/eldds1400c5-dataset BELI 593 https：//github.com/TheMakiran/BenchmarkELimages PVEL-AD 36 543 http：//aihebut.com/col.jsp？id=118 PVMD 1 108 https：//github.com/CCNUZFW/PV-Multi-Defect ISM 20 000 https：//github.com/RaptorMaps/InfraredSolarModules CS 1 837 https：//datahub.duramat.org/dataset/crack-segmentation
Table 8. Evaluation metrics of classification and segmentation
View tableView in ArticleTable 8. Evaluation metrics of classification and segmentation
Metric Formula Applicability Accuracy↑ Classification
Segmentation
Recall↑ Precision↑ F1-score↑ FPR↓ FNR↓ IoU↑ Segmentation MIoU↑ Dice coefficient↑
Table 9. Performance comparison and analysis of existing traditional machine vision algorithms
View tableView in ArticleTable 9. Performance comparison and analysis of existing traditional machine vision algorithms
Method Detection object Detection approach Sample Size Performance Feature+RBF-SVM［ 34 ］Lack of angle， Broken gate， Collapse edge， crack， Leaking slurry， Casting point Classification 750 defect samples Recognition rate： 90% and above KAZE+SVM［ 35 ］likelihood of defects Classification 2624 images， Size：
300×300 pixels
Accuracy 88.24% SVM［ 39 ］Parallel cracks， Series cracks， Bypass diode failure， Defect-free， Random defective solar cells. Classification 100 defect-free images，
100 images with defects，
Pixel value： uint8
Accuracy： 97.5% RBF+SVM［ 40 ］Appearance defects， Color defects， Cracks defects. Classification 600 defect samples， Size：
656×492 pixels
Recognition rate： 98.57%， Computation time 0.107 s DBF+SVM［ 41 ］likelihood of defects Classification 2624 images， Size：
300×300 pixels
4 classes： 90.57%
2 classes： 94.52%
Haar feature +clustering［ 45 ］Micro-crack， break， finger-interruption. Pixel-level Segmentation 31 defect-free images，
19 images with defects
Size： 550×550 pixels
Processing time per image 0.1 s Low-rank representation［ 51 ］Surface defect Pixel-level Segmentation 40 images with defects，
60 images without defects， Size： 226×226 pixels
Precision： 0.867
F1-score： 0.778
Fourier transform［ 54 ］Micro-crack， break， finger-interruption. Pixel-level Segmentation Image size 550×550 pixels divided into 323 images of size 75×75， 308 defect-free samples， 15 samples with defects Recognition rate： 1.00 Speed： 0.006 s per sub-image 0.29 s per original image Binary and Discrete Fourier Transform［ 55 ］Micro-crack Classification + Detection Window 60 EL benchmark images， Size： 720×720 pixels Detection time for 60 cells
High resolution： 1.62 s Low resolution： 2.52 s
ICA［ 58 ］Internal defect Pixel-level Segmentation 28 defect-free images， 52 defective images， Size： 1 250×1 250 pixels 0.04 seconds per image
Table 10. Comparison of the performance of existing representative algorithms on ELPV datasets
View tableView in ArticleTable 10. Comparison of the performance of existing representative algorithms on ELPV datasets
Method Accuracy Precision Recall F1-score Device SVM-KAZE［ 35 ］82.44% — — 82.52% Intel i7-3770 K CPU 32 GB RAM AC-VGG-19［ 35 ］88.42% — — 88.89% Two NVIDIA GeForce GTX 1080 DFB-SVM［ 41 ］89.63% 87.01% 91.58% 91.55% Intel Core i5 4200U 8 GB RAM L-CNN［ 41 ］82.58% 80.71% 96.59% 85.80% Intel Xeon E-2224G 16 GB RAM VGG-11（Lightweight）［ 66 ］93.02% 92.50% 92.00% 92.49% Intel Core i5，3.20 GHz CPU Inception-ResNet-v2［ 69 ］93.00% — — — 1 NVIDIA TITAN GPU ADI［ 71 ］83.00% — — — 4 NVIDIA TITAN GPUs HFCNN9［ 75 ］88.38% 88.00% 76.00% 82.00% 16 GB内存的NVIDIA TESLA P100 GPU RBF-SVM-SURF［ 77 ］72.74% — — — Intel Xeon E5-2680 v4，2.4 GHz CPU Improved CNN［ 77 ］91.58% 91.57% 91.85% 91.57% Nvidia GeForce GTX 1080 Ti SeF-HRNet［ 78 ］94.90% 93.85% 94.29% 94.07% RTX 3090 GPU PSO Pruner-CNN［ 80 ］90.91% — — 94.16% NVIDIA GeForce RTX 3080 VGG-16+SVM［ 82 ］99.49% — — — — DPiT［ 95 ］91.70% 93.20% 79.60% 85.90% 8 GeForce GTX 1080Ti GPUs Improved VGG-16［ 103 ］95.20% — 95.40% 94.40% NVIDIA TITAN GPU EfficientNet-B1［ 104 ］84.00% — — — 16 GB GPU Tesla T4 NAS+KD［ 108 ］91.74% 87.13% 86.28% 86.70% Intel i9-10980XE 24.85 M Improved ResNet50-L1［ 116 ］94.00% 90.00% 77.00% 83.00% —
Tools
Get Citation
Copy Citation Text
Yuqi LIU, Yiquan WU. Review of defect detection algorithms for solar cells based on machine vision[J]. Optics and Precision Engineering, 2024, 32(6): 868
Paper Information
Category:
Received: Sep. 25, 2023
Accepted: --
Published Online: Apr. 19, 2024
The Author Email: WU Yiquan (nuaaimage@163. com)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20