SPACE: SPAtial-aware Consistency rEgularization for anomaly detection in Industrial applications

TL;DR

Experimental results, through qualitative evaluation, demonstrate the superiority of detection and efficiency of each module compared to state-of-the-art methods.

Abstract

In this paper, we propose SPACE, a novel anomaly detection methodology that integrates a Feature Encoder (FE) into the structure of the Student-Teacher method. The proposed method has two key elements: Spatial Consistency regularization Loss (SCL) and Feature converter Module (FM). SCL prevents overfitting in student models by avoiding excessive imitation of the teacher model. Simultaneously, it facilitates the expansion of normal data features by steering clear of abnormal areas generated through data augmentation. This dual functionality ensures a robust boundary between normal and abnormal data. The FM prevents the learning of ambiguous information from the FE. This protects the learned features and enables more effective detection of structural and logical anomalies. Through these elements, SPACE is available to minimize the influence of the FE while integrating various data augmentations.In this study, we evaluated the proposed method on the MVTec LOCO, MVTec AD, and VisA datasets. Experimental results, through qualitative evaluation, demonstrate the superiority of detection and efficiency of each module compared to state-of-the-art methods.

Figures (10)

• Figure 1: The ambiguities of applying strong augmentation for industrial data.
• Figure 2: Comparison of existing and proposed approaches for anomaly detection: (a) represents the basic S-T structure, (b) shows the structure combining S-T and FE, and (c) depicts the SPACE structure proposed.
• Figure 3: The overall architecture: The method combines a structural branch, detecting fine-grained anomalies through consistency regularization, and a logical branch, focusing on shape anomalies using the FM.
• Figure 4: The region for updating features in augmented images: In (c) and (e), the color scale signifies that a shift toward red corresponds to a larger number of channels employed for learning in the feature, while a shift toward blue indicates the use of fewer channels.
• Figure 5: Differences in structural anomaly maps when using FM: From left to right, the sequence represents the original image, the ground truth mask, the anomaly detection map without FM, and the anomaly detection map with FM.