Fig 1.
The defect detection system for solar cells.
Notes: 1-conveyer belt, 2-the light source of WS4, 3-the camera of WS4, 4-the camera of WS3, 5-the telecentric, 6-the light source of WS3, 7-the camera of WS2, 8-the light source of WS2, 9-the camera of WS1, 10-the light source of WS1, 11-solar cell.
Fig 2.
The field diagram of the image detection system.
Fig 3.
The 5 types defects of solar cells.
Notes: (a) mismatch defect, (b) bubble defects, (c) cell-crack defects, (d) glass-crack defect, (e) glass-upside-down defect 1, (f) glass-upside-down defect 2.
Table 1.
The defect datasets before and after image enhancement.
Fig 4.
The structure of YOLOv5s model.
Fig 5.
The steps of K-means algorithm and the visualization of the clustering result of mismatch defects.
(a) The process of obtaining the optimal size of the anchor boxes by the K-means algorithm. (b) The visualization results of the clustering process for mismatch defects labeled boxes.
Fig 6.
The comparison of 9 anchor boxes before and after clustered calculation.
(a) The 9 anchor boxes before K-means clustered calculation. (b) The 9 anchor boxes after K-means clustered calculation.
Fig 7.
The loss and AP curves of default and improved anchor boxes.
(a) The loss curves. (b) The AP curves.
Table 2.
The detection accuracy indicators of the YOLOv5s model.
Fig 8.
Test results of mismatch defects before and after the anchor boxes are improved.
(a) The detection results of original YOLOv5 model. (b) The detection results of YOLOv5 model with improved anchor boxes.
Fig 9.
The structures of coupled head and decoupled head.
Fig 10.
The variation process of various loss functions before and after model improvement.
(a) Loss curves for predicting category information, Cls_loss. (b) Loss curves for prediction boxes information, Reg_loss. (c) Loss curves for predicting the inclusion or exclusion of objects, Obj_loss.
Fig 11.
The testing effectiveness comparison for general defects before and after the model is improved.
(a) The detection effectiveness comparison for cell-crack defects. (b) The detection effectiveness comparison for bubble defects. (c) The detection effectiveness comparison for glass-crack defects.
Table 3.
Comparative experiments of general defect detection.
Table 4.
Detection results of three models for glass-upside-down defects.
Fig 12.
The classification detection scheme for solar cell defects.
Table 5.
Detection result statistics of the classification detection scheme.