EntroAD: Structural Entropy-Guided Prompt Adaptation for Zero-Shot Anomaly Detection

Problem
This paper addresses the limitations of existing zero-shot anomaly detection (ZSAD) methods, particularly those based on CLIP, which typically utilize a single adaptation pathway. This approach is inadequate for handling the heterogeneous nature of anomalies that can vary significantly in characteristics, from prominent structural disruptions to subtle irregularities. The authors propose a novel framework, EntroAD, to enhance the adaptability of ZSAD systems across diverse domains without requiring target-domain adaptation. This work is presented as a preprint and has not yet undergone peer review.

Method
EntroAD introduces a structural entropy-guided approach to ZSAD, leveraging a dynamic routing mechanism that processes various anomaly types through specialized adaptation strategies. The core technical contributions include:

1. Structural Entropy Estimation: The framework computes patch-level structural entropy based on self-attention-induced relations among image patches. This entropy serves as a proxy for relational uncertainty, guiding the routing of tokens that are sensitive to different anomaly characteristics.

2. Dynamic Token Routing: Anomaly-aware routed tokens are constructed using the entropy estimates, allowing the model to focus on capturing distinct anomaly cues effectively.

3. Dual-Branch Prompt Adaptation Module: This module stabilizes visual-text alignment while maintaining the transferable prior knowledge of CLIP. It operates in a confidence-aware manner, enhancing the robustness of the model against variations in anomaly presentation.

The authors do not disclose specific details regarding the training compute or the exact architecture used, but the methodology emphasizes the integration of structural entropy into the anomaly detection pipeline.

Results
EntroAD demonstrates superior performance across 10 industrial and medical benchmarks in challenging cross-dataset ZSAD scenarios. The framework achieves state-of-the-art results, outperforming existing CLIP-based methods. While specific numerical results are not provided in the abstract, the authors claim significant improvements over named baselines, indicating a robust effect size in anomaly detection efficacy.

Limitations
The authors acknowledge that their approach may still struggle with extreme outlier cases or highly complex anomaly patterns that deviate significantly from the training data. Additionally, the reliance on self-attention mechanisms may introduce computational overhead, which could limit scalability in real-time applications. The paper does not address potential biases in the training datasets or the generalizability of the model to entirely novel anomaly types.

Why it matters
EntroAD’s introduction of structural entropy as a guiding principle for token routing in ZSAD represents a significant advancement in the field, particularly for applications requiring robust anomaly detection across diverse domains. The framework’s ability to adaptively process different anomaly types could lead to improved performance in critical areas such as industrial monitoring and medical imaging, where timely and accurate anomaly detection is essential. This work lays the groundwork for future research into multi-faceted anomaly detection strategies that can leverage structural insights for enhanced interpretability and effectiveness.

Authors: Xinyu Zhao, Qingyun Sun, Jiayi Luo, Jianxin Li
Source: arXiv:2605.28630
URL: https://arxiv.org/abs/2605.28630v1