Answers are grounded in this paper and cite their sources.
This paper addresses the challenge of multi-class unsupervised visual anomaly detection in unconstrained environments, aiming to enhance the robustness of existing frameworks against complex and uncontrolled background variations. While current anomaly detection approaches excel in controlled industrial settings, they often struggle in open and cluttered scenes, such as power-transmission facility inspection, where diverse background elements are difficult to model and are frequently misclassified as anomalies. Conventional solutions that rely on auxiliary segmentation models typically incur substantial computational overhead and introduce heavy dependencies on external systems, limiting their practicality in real-world deployments. To overcome these limitations, we propose a novel framework that leverages the intrinsic foreground-background discriminability of pre-trained DINO encoders to inject semantic priors directly into the representation learning process. Specifically, we integrate a lightweight classification head to the frozen encoder to efficiently extract foreground priors while introducing negligible inference latency. These priors are then used to guide three complementary mechanisms that explicitly decouple foreground modeling from background interference. First, we introduce a foreground-aware probabilistic masking strategy at the bottleneck, encouraging the model to allocate its reconstruction capacity to foreground semantics. Second, we impose an inter-region attention constraint within the decoder to explicitly penalize attention weights directed from foreground tokens to background distractors. Third, we propose a foreground-centric hard-patch reconstruction objective that prioritizes difficult foreground regions during optimization while dynamically suppressing background-induced anomaly scores. Extensive experiments on multiple benchmarks demonstrate that our method significantly improves detection accuracy and robustness in complex outdoor scenarios, offering a computationally efficient solution for complex real-world scenarios.
No takes yet. Share an insight, caveat, or question.
E et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: