Fig. 1. EL images of photovoltaic cells with different types of defects. (a) Normal; (b) crack; (c) star crack; (d) finger; (e) black core; (f) horizontal dislocation; (g) vertical dislocation; (h) short circuit

Fig. 2. EL image preprocessing of photovoltaic cells. (a) Original image; (b) (c) image augmentation; (d) image enhancement

Fig. 3. Dataset information. (a) Number of labels for each type of defect; (b) size and number of target boxes

Fig. 4. Flow chart of photovoltaic cell defect detection

Fig. 5. Three-stage cascade lightweight network structure

Fig. 6. Structure diagram of the CFB module

Fig. 7. EMA attention mechanism

Fig. 8. Structure diagram of EMSHead

Fig. 9. EMSConv lightweight convolution

Fig. 10. Loss curve and mAP variation curve during training process

Fig. 11. PR curves of the model in test set

Fig. 12. Multi-target defect detection results of the YOLO-FEE model. (a)(c) Multi-target detection results; (b) thermal map of Fig.12(a); (d) thermal map of Fig.12(c)

Fig. 13. Single target defect detection results of the YOLO-FEE model. (a) Crack; (b) star crack; (c) finger; (d) black core; (e) horizontal dislocation; (f) vertical dislocation; (g) short circuit