Fig. 1. Images of subpixel. (a) (b) Origin image; (c) (e) partial enlarged image of defect; (d) (f) partial enlarged image of normal area

Fig. 2. Images of significance detection’s spatial information. (a) Single scale; (b) multiple scales

Fig. 3. Saliency detection results of small size defect at different scales. (a) Origin image; (b) image with scale 1; (c) image with scale 2; (d) image with scale 3

Fig. 4. Saliency detection results of large size defect. (a) Origin image; (b) image with scale 1; (c) image with scale 2; (d) image with scale 3; (e) 2-scale fusion; (f) 3-scale fusion

Fig. 5. Result of defect region binarization. (a) Origin image; (b) multiscale saliency detection; (c) binarization; (d) interference cancellation

Fig. 6. Illustration of complete defect area acquisition process. (a) Binarization; (b) image of defect and gaps; (c) convex hull fitting; (d) final result

Fig. 7. Illustration of defect block grouping. (a) Origin image; (b) image of defect and gaps; (c) equality of both sides; (d) inequality of both sides

Fig. 8. Illustration of defect block connection. (a) Local convex hull fitting; (b) final result

Fig. 9. Comparison of segmentation results by different methods. (a) Origin image; (b) manual segmentation; (c) proposed method; (d) contrast experiment 1; (e) contrast experiment 2