Separating Novel Features for Logical Anomaly Detection: A Straightforward yet Effective Approach

TL;DR

This work tackles false negatives in knowledge distillation–based logical anomaly detection for industrial vision. It introduces Distinctive Feature Separation Constraint (DFSC), a training constraint added to EfficientAD that preserves a discriminative gap between the teacher and auto-encoder representations to better capture novel features indicative of logical anomalies. Evaluated on MVTec LOCO AD, the approach achieves an average AUROC improvement of about 1.3% and improved localization, demonstrating that a simple, unsupervised constraint can enhance logical anomaly detection without major architectural changes. The method promises more reliable, real-time anomaly detection in manufacturing by mitigating false negatives in global feature learning.

Abstract

Vision-based inspection algorithms have significantly contributed to quality control in industrial settings, particularly in addressing structural defects like dent and contamination which are prevalent in mass production. Extensive research efforts have led to the development of related benchmarks such as MVTec AD (Bergmann et al., 2019). However, in industrial settings, there can be instances of logical defects, where acceptable items are found in unsuitable locations or product pairs do not match as expected. Recent methods tackling logical defects effectively employ knowledge distillation to generate difference maps. Knowledge distillation (KD) is used to learn normal data distribution in unsupervised manner. Despite their effectiveness, these methods often overlook the potential false negatives. Excessive similarity between the teacher network and student network can hinder the generation of a suitable difference map for logical anomaly detection. This technical report provides insights on handling potential false negatives by utilizing a simple constraint in KD-based logical anomaly detection methods. We select EfficientAD as a state-of-the-art baseline and apply a margin-based constraint to its unsupervised learning scheme. Applying this constraint, we can improve the AUROC for MVTec LOCO AD by 1.3 %.

Figures (4)

• Figure 1: It illustrates the potential false negative problem that knowledge distillation-based methods such as EfficientAD batzner2023efficientad can experience on some samples. (c) indicates that a bolt must be somewhere in the white area. If there is no appropriate constraint during unsupervised learning, the difference map cannot represent meaningful difference. Accordingly, given an image which has permissible objects containing logically abnormal patterns such as (a), no abnormal pattern may be capture as (b). On the other hand, when applying the proposed constraint, the existence of an logically abnormal pattern can be detected to some extent as (d).
• Figure 2: The overview of the proposed constraint. At training time, only normal image is fed into teacher and auto-encoder. We split output features from teacher and auto-encoder by the quantile value $\tau^{'}$ into similar and novel features. Then, we enforce auto-encoder to keep far distance by margin for novel features. Global representation is obtained from channel mean of difference map between $S^{A}$ and $A$. Since $S^{A}$ is trained on $d^{SA}(x)$ in Eq. \ref{['eq1']}, when an image is given, $S^{A}$ and $A$ produce different features on novel pattern. Consequently, our constraint leads $S^{A}$ and $A$ to keep distance on the novel pattern.
• Figure 3: (a) and (b) represent input images and ground truth masks for logical anomalies. If a model captures logical anomalies in the white areas of (b), those are considered as correct. (c) is the predicted anomaly maps produced by EfficientAD-S denoted as EAD-S. (d) represents the predicted anomaly maps produced by our method.
• Figure 4: Performance variations versus quantile values for MVTec LOCO AD. Evaluation metric is image-level AUROC (%). The legends mean the quantile values for our constraint.