SoftPatch: Unsupervised Anomaly Detection with Noisy Data

• DOI: 10.48550/arxiv.2403.14233
• Publication Date: 2024-03-21
• Authors: Xi Jiang; Huiling Chen; Qiang Nie; Yong Liu; Jianlin Liu; Bin-Bin Gao; Jun Liu; Chengjie Wang; Feng Zheng
• Keywords: Anomaly (physics)

Although mainstream unsupervised anomaly detection (AD) algorithms perform well in academic datasets, their performance is limited practical application due to the ideal experimental setting of clean training data. Training with noisy data an inevitable problem real-world but seldom discussed. This paper considers label-level noise image sensory for first time. To solve this problem, we proposed a memory-based AD method, SoftPatch, which efficiently denoises at patch level. Noise discriminators are utilized generate outlier scores patch-level elimination before coreset construction. The then stored memory bank soften boundary. Compared existing methods, SoftPatch maintains strong modeling ability normal and alleviates overconfidence coreset. Comprehensive experiments various scenes demonstrate that outperforms state-of-the-art methods on MVTecAD BTAD benchmarks comparable those under without noise.