EEE-Net: Efficient Edge Enhanced Network for Surface Defect Detection of Glass

Surface defect detection is essential for ensuring product quality. Intelligent defect detection is widely studied and applied in many industrial fields. However, glass defect detection is a daunting task because the optical properties of glass present unique challenges, e.g., intra-class difference, low contrast, and ambiguous edges. In this paper, we propose an efficient edge enhancement network (EEE-Net) to address the above challenges. EEE-Net employs Efficient Transformer Blocks to compose the pyramid network; each block combines a sequence reduction block (SRB) for efficient long-range contextual modeling and feature refinement. We propose three modules for edge enhancement: an encoder for edge feature (En-edge), a decoder for edge feature (De-edge) and an edge information fusion module (EIF). En-edge and De-edge are encoders and decoders of edge features. They extract and enhance the edge information of the network at each layer, respectively, focusing on the edge change points of the network. The EIF module is used to fuse multiple layers of feature and edge information, enabling the network to obtain accurate defect outlines. Experimental results on Glass Surface Defect (GSD) and Mobile phone screen Surface Defect (MSD) datasets show the superiority of the proposed model and its feasibility for real-time industrial applications.