Fig. 1. YOLOv5s network structure

Fig. 2. Structure of GhostV2 bottleneck module. (a) Bottleneck structure when the convolution stride is 2; (b) bottleneck structure when the convolution stride is 1

Fig. 3. Structure of C3GhostV2 module

Fig. 4. Structure of ODConv module

Fig. 5. Structure of improved YOLOv5s

Fig. 6. Dataset categories for laser soldering defects. (a) Normal; (b) less tin; (c) poly tin; (d) fired tin; (e) unwelded tin; (f) no tin; (g) continuous tin; (h) soldering through

Fig. 7. Sample distribution of laser soldering solder joint defect dataset before and after expansion

Fig. 8. Comparison of experimental effects of YOLOv5s before and after improvement on self-made dataset. (a) YOLOv5s; (b) improved YOLOv5s

Fig. 9. mAP@0.5 training curves of YOLOv5s and its different improved versions

Fig. 10. Thermodynamic charts of fired tin solder joint defect characteristic before and after neck network improvement

Fig. 11. Simulated experiment of laser soldering solder joint defect detection