Fig. 1. Physical drawing of the welding equipment

Fig. 2. SCUNet network structure

Fig. 3. SC Block structure

Fig. 4. MixFormer network architecture

Fig. 5. Schematic diagram of Mask2Former structure

Fig. 6. Knowledge distillation schematic

Fig. 7. Denoising results of different algorithms. (a) Original image; (b) SCUNet; (c) median filtering; (d) mean filtering; (e) Gaussian filtering

Fig. 8. Tracking results comparison of different algorithms. (a) MixFormer; (b) STARK; (c) SiamRPN++

Fig. 9. Variation in evaluation indicators for the backbone network

Fig. 10. Variation in backbone network loss

Fig. 11. Variation in knowledge distillation evaluation indicator

Fig. 12. Change curves of segmented network evaluation indicator

Fig. 13. Ten experimental results of reasoning speed of split networks

Fig. 14. Segmentation results of different algorithms. (a)(d) Laser melting; (b) laser wire-filling welding; (c) laser molten pool under high-frequency vibration; (e) laser powder-feeding welding; (a₁)‒(e₁) segmentation result of DeepLabV3+ network; (a₂)‒(e₂) segmentation results of FCN network; (a₃)‒(e₃) segmentation results of PSPNet network; (a₄)‒(e₄) segmentation results of Mask2Former network; (a₅)‒(e₅) segmentation results of the proposed algorithm