Fig. 1. Schematic of weld information measurement

Fig. 2. Projection model measurement (Oc-XcYcZc and uv are camera coordinate system and image coordinate system, respectively)

Fig. 3. Network structure. (a) Our network; (b)-(d) Detail/Seg Head, ARM, and FFM modules used in the model

Fig. 4. STDC module

Fig. 5. DSNT module

Fig. 6. Training loss of proposed model

Fig. 7. Detail label generation

Fig. 8. Laser stripe segmentation results. (a) Original images; (b) laser stripe label images; (c) weld seam features extracted by FCN-8s; (d) features extracted by our method with detailed information supervision; (e) features extracted by our method without detailed information supervision

Fig. 9. Location results of weld feature points by DSNT method under different noise interferences, where the green and blue “+” are left and right feature points and yellow “+” is intermediate feature point

Fig. 10. Comparison of feature point location results. (a1)-(a3) Extraction errors of left feature point, intermediate feature point, and right feature point in u-axis direction in weld image, respectively; (b1)-(b3) left feature point, intermediate feature point, and right feature point in v-axis direction in weld image, respectively

Fig. 11. Comparison of location errors of subtask correlation feature points. (a) In u-axis direction; (b) in v-axis direction

Fig. 12. Time consumption in processing for image sequence

Table 1. Feature point positioning branch structure

Table 2. Comparison of laser stripe segmentation accuracy

Table 3. Comparison of inference time of different networks