Deep Feature Contrasting for Industrial Image Anomaly Segmentation

Industrial image anomaly segmentation is pivotal in ensuring the quality inspection of products within intelligent manufacturing systems. Recent research efforts have predominantly focused on deep learning-based approaches to address this challenge. However, unsupervised methods are often susceptible to distribution shifting, while supervised methods face significant obstacles due to imbalanced samples, resulting in suboptimal accuracy for anomaly segmentation. This article introduces a novel end-to-end deep feature contrasting (DFC) method aimed at enhancing segmentation accuracy in scenarios with limited supervision and pixel-level labeled anomalous images. DFC includes the introduction of a novel backbone feature contrasting pyramid (BFCP) featuring dual-model channels, which effectively captures the distribution of normality. In addition, a novel model-independent semantic feature contrasting (SFC) technique, implemented through self-contrasting, is proposed to train a discriminative segmenting head module (SHM) that addresses the challenge of sample unbalance. Furthermore, this article presents two types of SFC, namely, global and local, which are suggested to improve the overall performance. The proposed DFC method achieves F1 of 0.787 and 0.665 on two real-world industrial datasets (Kolektor Surface Defect Dataset (KSDD) and KSDD2) and is applied to a practical application of anomaly segmentation of printed circuit boards, demonstrating the superior anomaly segmenting performance.