A Hierarchical Patch Feature Distribution Network for Industrial Multiscale Defect Detection

Visual inspection of surface defects in industrial products is crucial for quality control but remains challenging due to unpredictable and multiscale defects. Unsupervised anomaly detection methods based on feature distribution have rapidly developed, requiring only normal samples and no additional manual labeling. However, these methods still suffer from feature bias and image misalignment, leading to overdetection of multiscale defects, making them impractical for deployment in factory settings. To address these issues, this study proposes a hierarchical patch feature distribution modeling (HPDM) network. Specifically, the new patch feature alignment learning (PFAL) module mitigates image misalignment by enhancing feature dimension similarity via the PFAL loss function. Additionally, the proposed discrete feature memory (DFM) module employs linear mapping to increase the density of the normal feature distribution, addressing feature bias from pretrained networks. Furthermore, the pyramid-shaped Gaussian distribution (PGD) module fits multivariate Gaussian models at different feature scales for defect detection across various scales. On the MVTec2D, MVTec3D, and VISION datasets, the HPDM network achieves pixel-level AUROC scores of 98.83%, 98.30%, and 71.20%, respectively, surpassing those of the existing methods and demonstrating state-of-the-art anomaly detection performance. The experimental results on a real printed circuit board assembly (PCBA) circuit dataset demonstrate a detection speed of 30 frames per second (FPS), meeting the requirements for factory online detection. Note to Practitioners—Previous unsupervised anomaly detection methods have struggled to detect both large and small surface defects simultaneously. However, the proposed multiscale patch feature modeling network effectively identifies various multiscale defects on product surfaces, including PCBA, thin-film transistor liquid crystal displays, and tiles. Additionally, HPDM requires only a small number of normal samples to learn a robust network, which is crucial for industrial applications where identifying and labeling samples is challenging. Furthermore, by leveraging model pruning and quantization techniques, HPDM can be applied to online visual inspection.