Research on image segmentation method of steel slag layer in RH vacuum degassing

Abstract

In the RH vacuum degassing refining process, the measurement of the slag layer thickness is crucial. This paper provides a light-weight improved model on the basis of DeepLabv3+. Firstly, MobileNetV2 is used as the backbone network for feature extraction. This improves training efficiency and reduces model complexity. Secondly, regular convolutions in the ASPP module are replaced by depth-separable convolutions to increase the speed of the calculation. Finally, efficient multi-scale attention is introduced into high-level features to enhance segmentation accuracy. The experimental results show that, in comparison with the original DeepLabv3+, the, the proposed method reduces model parameters by 89.37%, increases mIoU by 1.89%, and balances model complexity and accuracy, demonstrating high practicality.