SoftPatch: Unsupervised Anomaly Detection with Noisy Data

TL;DR

Comprehensive experiments demonstrate that SoftPatch outperforms the state-of-the-art AD methods on the MVTecAD and BTAD benchmarks and is comparable to those methods under the setting without noise.

Abstract

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

Figures (8)

• Figure 1: Illustration of SoftPatch. Unlike previous methods that construct coreset without considering the negative effect of noisy data, SoftPatch wipes off easy noisy data to formulate a clean training set and alleviates hard noisy data's impact by soft-reweighting.
• Figure 2: Overview of the proposed method. In the training phase, the noises are distinguished at patch level at each position of the feature map by a noise discriminator. The deeper color a patch node has, the higher probability that it is a noise patch. After achieving outlier scores for all patches, the top $\tau$% patches with the highest outlier score are removed. The coreset is a subset of remaining patches after denoising. Different from other methods, our memory bank consists of the samples in coreset and their outlier scores which are stored as soft weights. Soft weights will be further utilized to re-weight the anomaly score in inference.
• Figure 3: The comparison of anomaly detection performance under noisy training. no overlap means the injected anomalous images are removed from test set while overlap are not.
• Figure 4: Performance trend with the threshold $\tau$ in SoftPatch-LOF. The results are evaluated on MVTecAD-noise-0.1.
• Figure 5: Comparison between corsets of AD methods with same noisy train set, MVTecAD-Pill with noise-0.1. We use t-SNE for dimension reduction for visualization. The yellow dots represent patch features from noisy sample, while the purple dots are nominal. Compared with the other two, SoftPatch wipe off the noisy patch and model the nominal data properly.