Unsupervised Defect Segmentation via Forgetting-Inputting-Based Feature Fusion and Multiple Hierarchical Feature Difference

In the field of surface defect detection, there is a significant imbalance between the number of positive and negative samples, which has led to a growing interest in positive-samples-based anomaly detection methods. Reconstruction-based methods are currently the most commonly used approach, but they often struggle to repair abnormal foregrounds and reconstruct clear backgrounds simultaneously. To address this issue, we propose a new approach called the forgetting-inputting-based feature fusion and multiple hierarchical feature difference network (FIM-Net). The FIM-Net method incorporates a novel latent feature repair module (LFRM), which combines encoding and memory-encoding obtained by memory-augmented module (MAM) via a novel forgetting-inputting-based feature fusion module (FIFFM) to repair abnormal foregrounds while preserving clear backgrounds. Additionally, we introduce a manual defect generation algorithm (MDGA) to simulate realistic and feature-rich anomalies. Finally, we use a multiple hierarchical feature difference (MHFD) for defect segmentation to achieve more accurate defect location. Our extensive comparison experiments demonstrate that the FIM-Net method achieves the state-of-the-art detection accuracy and shows great potential for industrial applications.