A Coarse and Fine Grained Network for Industrial Surface Defect Classification

Automated surface defect detection plays a critical role in maintaining product quality and ensuring high standards in industrial settings. In this work, we present a novel approach for surface defect detection in industrial settings using a coarse and fine grained (CFG) network built upon the Vision Transformers (ViT) framework. Our CFG network consists of two branches: a coarse-grained branch and a fine-grained branch. The coarse-grained branch focuses on capturing the overall structure and layout of the inspected surfaces, facilitating a high-level comprehension of objects and their spatial arrangement. In contrast, the fine-grained branch is designed to extract intricate local details and subtle irregularities that could indicate the presence of defects. To enhance performance, we integrate a cross fusion module that allows for adaptive fusion of the coarse and fine-grained representations. This adaptive fusion enables the network to dynamically prioritize either global or local features depending on the characteristics of the input image. Experimental validation using the NEU-CLS dataset demonstrates competitive improvements in performance compared to existing approaches.