Global Information Attention Based Dual-Pathway Network for Oxidized Slag Segmentation of Metal Ingot Images

Abstract

The content of oxidized slag on the surface of metal ingot is the key to judge whether the metal ingot is qualified. However, it is still a challenge to determine the covering area of oxidized slag on the surface of metal ingot. In this paper, we proposed a novel Global Information Attention based dual-pathway Network (GANet) consists of context and edge pathways to segment oxidized slag on the metal ingot surface image. Specifically, to suppress the interference of background on oxidized slag segmentation, we propose novel Fusion Attention Module (FAM) and Global Position Attention Module (GPA), which can effectively aggregate global semantic information and improve segmentation performance. Meanwhile, we designed the edge pathway and Edge Perception Module (EPM) to extract the boundary information of the oxidized slag. The experimental results show that the performance of the GANet is better than the existing segmentation methods which proves the effectiveness of the proposed method and provides a research idea for the automation of metal casting process.