Fig. 1. Different forms of defects in PV modules. (a) Micro-crack in polycrystalline silicon; (b) micro-crack in monocrystalline silicon; (c) finger-interruption; (d) break and black-zone

Fig. 2. Experimental results of EL image segmentation. (a) EL image of a PV module; (b) EL subimage in row 1 and column 5

Fig. 3. Network structure of the improved YOLOv5 algorithm

Fig. 4. Schematic of Ghost convolution module

Fig. 5. Schematic of the SE attention convolution module

Fig. 6. Schematic of BiFPN

Fig. 7. Trend chart of results during training

Fig. 8. Comparison of the detection performance for different kinds of defects. (a) Confusion matrix; (b) P-R curve

Fig. 9. Visualisation of the highest-level feature maps by class activation mapping (grad-CAM) before and after algorithm improvement

Fig. 10. Visual presentation of partial test results