Explicit Boundary Guided Semi-Push-Pull Contrastive Learning for Supervised Anomaly Detection (2207.01463v2)

Abstract: Most anomaly detection (AD) models are learned using only normal samples in an unsupervised way, which may result in ambiguous decision boundary and insufficient discriminability. In fact, a few anomaly samples are often available in real-world applications, the valuable knowledge of known anomalies should also be effectively exploited. However, utilizing a few known anomalies during training may cause another issue that the model may be biased by those known anomalies and fail to generalize to unseen anomalies. In this paper, we tackle supervised anomaly detection, i.e., we learn AD models using a few available anomalies with the objective to detect both the seen and unseen anomalies. We propose a novel explicit boundary guided semi-push-pull contrastive learning mechanism, which can enhance model's discriminability while mitigating the bias issue. Our approach is based on two core designs: First, we find an explicit and compact separating boundary as the guidance for further feature learning. As the boundary only relies on the normal feature distribution, the bias problem caused by a few known anomalies can be alleviated. Second, a boundary guided semi-push-pull loss is developed to only pull the normal features together while pushing the abnormal features apart from the separating boundary beyond a certain margin region. In this way, our model can form a more explicit and discriminative decision boundary to distinguish known and also unseen anomalies from normal samples more effectively. Code will be available at https://github.com/xcyao00/BGAD.

Explicit Boundary Guided Semi-Push-Pull Contrastive Learning for Supervised Anomaly Detection

The paper introduces a novel approach to enhancing supervised anomaly detection (AD) using a method termed Explicit Boundary Guided Semi-Push-Pull Contrastive Learning. Traditional AD models predominantly rely on unsupervised techniques, focusing solely on normal samples to delineate anomaly characteristics. However, this strategy inherently suffers from ambiguous decision boundaries due to a lack of direct exposure to anomaly samples during the training phase. Recognizing the availability of partial anomaly data in practical scenarios, the authors propose an innovative approach aimed at effectively exploiting these known anomalies while addressing the bias issues they typically introduce.

Key Contributions

1. Explicit Boundary Generation: The authors employ a conditional normalizing flow to model the distribution of normal features robustly. This enables the determination of a compact and explicit separating boundary based solely on normal feature distributions, thereby alleviating bias concerns from known anomalies.
2. Boundary Guided Optimization: Introducing the boundary guided semi-push-pull (BG-SPP) loss, the approach uniquely integrates anomalies into the learning process without compromising the model's generalizability to unseen anomalies. Specifically, the BG-SPP loss functions to compress normal feature distributions and expand anomaly feature boundaries, thus establishing a clearer distinction between the two.
3. Generative Strategy for Anomalies: The paper also explores a RandAugment-based Pseudo Anomaly Generation, leveraging transformations to simulate anomalies and enhance the diversity and quantity of training data.

Results and Implications

The performance of the proposed method—dubbed BGAD—was rigorously evaluated across multiple datasets encompassing industrial defect inspection (e.g., MVTecAD and BTAD) and medical lesion detection (e.g., BrainMRI and HeadCT). BGAD exhibited superior results, consistently surpassing both unsupervised and other supervised AD methods in image-level and pixel-level AUROC metrics.

Implications:

• Practical Enhancements: This model demonstrates efficacy in real-world applications where both seen and unseen anomalies must be detected and characterized. Robust AD models like BGAD can significantly improve anomaly detection tasks in industries reliant on precision and reduction of false positives/negatives.
• Theoretical Pathways: The novel consideration of anomaly bias through explicit boundary guidance offers a new dimension to supervised learning paradigms, encouraging further exploration into other domains requiring anomaly detection.

Future Prospects

Potential advancements could focus on further refining the boundary selection mechanism, possibly integrating more adaptive or learning-based approaches to hyperparameter tuning. Moreover, exploration of alternative generative strategies for pseudo anomalies could yield greater variability and realism in training data, further enhancing model robustness.

This research not only contributes to anomaly detection methodologies but also offers a framework adaptable to various fields requiring precise differentiation between normal and anomalous data patterns. The explicit boundary and semi-push-pull mechanics present a promising direction for future studies in AI and machine learning.