Fig. 1. Image of cover window area and cropping diagram. (a) Cover window area diagram; (b) cropping diagram

Fig. 2. Cell phone cover part of the defect and labeling example diagrams. (a) Shadow scratch; (b) surface dust; (c) dirt; (d) black scratch; (e) white scratch; (f) white point; (g) black point

Fig. 3. Structure diagram of BiSeNet V2

Fig. 4. Example images of weighted image difference and feature enhancement. (a) Original defect image; (b) difference image; (c) difference overlay image

Fig. 5. Weighted image difference histogram. (a) Original defect histogram; (b) difference histogram; (c) difference overlay histogram

Fig. 6. Structure diagram of detail branch convolutional layer improvement. (a) Original convolutional block; (b) group dilation convolution block

Fig. 7. SE attention block

Fig. 8. SE attention embeded in semantic branching. (a) Stem block; (b) gather-and-expansion layer (GE 2)

Fig. 9. Gather-and-expansion layer and context embedding block. (a) Gather-and-expansion layer (GE 1); (b) context embedding block

Fig. 10. Bilateral guidance aggregation module

Fig. 11. Seg head block

Fig. 12. Multi-scale feature fusion decoding network

Fig. 13. Structure diagram of improved BiSeNet V2

Fig. 14. Comparisons of segmentation prediction results. (a) Original defect map; (b) defect labels; (c) original network segmentation prediction results; (d) improved network segmentation prediction results