Core Concepts
SoftPatchは、画像感覚の異常検出におけるノイズデータを考慮した効率的な手法です。
Abstract
Abstract:
Mainstream unsupervised anomaly detection algorithms perform well in academic datasets but struggle in practical applications due to noisy training data.
SoftPatch addresses label-level noise in image sensory anomaly detection for the first time.
Utilizes memory-based unsupervised AD method to denoise data at the patch level.
Introduction:
Unsupervised sensory anomaly detection is crucial for industrial applications where defects are hard to collect.
Existing methods rely on clean training sets, leading to performance degradation with noisy data.
Related Work:
Methods like agent tasks and knowledge distillation have been used for unsupervised anomaly detection.
Feature modeling directly models output features of the extractor for distribution estimation.
The Proposed Method:
SoftPatch introduces patch-level denoising strategy for coreset memory bank, improving data usage rate compared to sample-level denoising.
Noise discriminators and soft weights are utilized to enhance model robustness against noisy data.
Stats
本論文では、ノイズデータに対する画像感覚の異常検出において、SoftPatchが他の手法よりも優れた性能を示すことが示されています。