Image-Based Iron Slag Segmentation via Graph Convolutional Networks

Slagging-off (i.e., slag removal) is an important preprocessing operation of steel-making to improve the purity iron. Current manual-operated removal schemes are inefficient and labor-intensive. Automatic slagging-off desirable but challenging as reliable recognition iron difficult. This work focuses on realizing efficient accurate algorithm slag, which conducive realize automatic operation. Motivated by recent success deep learning techniques in smart manufacturing, we introduce methods this field for first time. The monotonous gray value industry images, poor image quality, nonrigid feature challenge existing fully convolutional networks (FCNs). To end, propose a novel spatial graph network (SFGCN) module. SFGCN module can be easily inserted FCNs reasoning ability global contextual information, helpful enhance segmentation accuracy small objects isolated areas. verify validity module, create industrial dataset conduct extensive experiments. Finally, results show that our brings consistent performance boost wide range FCNs. Moreover, adopting lightweight backbone, method achieves real-time segmentation. In future work, will dedicate efforts weakly supervised quick annotation big data stream generalization current models.