Novel industrial surface-defect detection using deep nested convolutional network with attention and guidance modules

Abstract Industrial defect inspection plays a crucial role in maintaining the high quality of the product. Although deep learning technologies have been applied to conduct automatic defect inspection, it is still difficult to detect industrial surface defects accurately due to complex variations. This study proposes a novel approach to industrial surface-defect detection that segments defect areas accurately and robustly from the complex background using a deep nested convolutional network (NC-Net) with attention and guidance modules. NC-Net consists of the encoder-decoder with nested residual U-blocks and feature enhancement modules. Each layer block of the encoder and decoder is also represented as a residual U-block. In addition, features are adaptively refined by applying the attention module to the skip connection between the encoder and decoder. Low-level encoder features are refined through edge guidance, and high-level encoder features through mask guidance, which can keep local and global contexts for accurate and robust defect detection. A comprehensive evaluation was conducted to verify the novelty and robustness of NC-Net using four datasets, including magnetic tile surface defects, steel surface defects, rail surface defects, and road surface defects. The proposed method outperformed previous state-of-the-art studies. An additional dataset was also evaluated to prove the extensibility and generality of the proposed approach.