Thickness measurement of immersion metal carbon slide based on image segmentation

Abstract

The thickness of a metal-immersed carbon slide mounted on a train’s flow shoe was measured by using machine vision and deep learning. A method for measuring the thickness of carbon slide plate based on improved U2-Net is proposed. Aiming at the problem that the edge feature extraction is not obvious, a new feature extraction module is designed. Efficient Channel Attention (ECA) mechanism and pool residual structure are used to make the network more suitable for metal-immersed carbon slide image segmentation. The experimental results show that the improved U2-Net network accuracy reaches 99,4 %, and the average absolute error is only 0,4 %. The thickness measurement accuracy of metallized carbon slide using improved U2-Net network reaches 0,5 mm.

Keywords

• rail
• immersion metal carbon slide
• image segmentation
• U²-Net