ONER: Online Experience Replay for Incremental Anomaly Detection

TL;DR

The paper tackles incremental anomaly detection in dynamic industrial settings, where catastrophic forgetting and cross-task feature conflicts hinder continual performance. It introduces ONER, an end-to-end online experience replay framework that combines decomposed prompts (learnable per-task components that reuse frozen past parameters) with semantic prototypes (image- and pixel-level feature banks) to preserve prior knowledge while adapting to new tasks. Through modules like Image-level Prototypes Refinement, Experience Replay-based Optimization, and Incremental Selection of Pixel-level Prototypes, ONER achieves state-of-the-art pixel-level and image-level anomaly detection on MVTec AD and VisA with very low training overhead (0.019M parameters and 5 epochs per task). The approach demonstrates strong scalability and practicality for real-world deployment, delivering robust discrimination and localization without revisiting old data, and enabling rapid adaptation to new product types and production schedules. Overall, ONER provides a data-efficient, stable solution for continual industrial anomaly detection that balances performance, efficiency, and memory usage.

Abstract

Incremental anomaly detection aims to sequentially identify defects in industrial product lines but suffers from catastrophic forgetting, primarily due to knowledge overwriting during parameter updates and feature conflicts between tasks. In this work, We propose ONER (ONline Experience Replay), an end-to-end framework that addresses these issues by synergistically integrating two types of experience: (1) decomposed prompts, which dynamically generate image-conditioned prompts from reusable modules to retain prior knowledge thus prevent knowledge overwriting, and (2) semantic prototypes, which enforce separability in latent feature spaces at pixel and image levels to mitigate cross-task feature conflicts. Extensive experiments demonstrate the superiority of ONER, achieving state-of-the-art performance with +4.4% Pixel AUROC and +28.3% Pixel AUPR improvements on the MVTec AD dataset over prior methods. Remarkably, ONER achieves this with only 0.019M parameters and 5 training epochs per task, confirming its efficiency and stability for real-world industrial deployment.

Figures (3)

• Figure 1: Comparison between a typical incremental AD method and ours: (a) Typical methods train task-specific prompts and require task identification during inference. (b) We employ an end-to-end prompt for both training and inference, bypassing the need for task identification. This avoids staged processing, minimizes cumulative errors, and enables seamless adaptation to new tasks while preserving prior knowledge by leveraging past experience effectively.
• Figure 2: This diagram illustrates the structure of ONER. Specifically, ONER incorporates two types of experience: decomposed prompts and semantic prototypes. Decomposed prompts, formed from learnable components, reuse previous knowledge to help the model learn new tasks. Semantic prototypes provide regularization at both pixel and image levels, preventing forgetting across tasks. The diagram shows the training process for step $t$, where both the decomposed prompts and semantic prototypes are updated after training.
• Figure 3: Qualitative evaluation on the MVTec AD dataset after training on the last subset. Experiments demonstrate our method's superior stability and knowledge retention over sample-replay (DRAEM-Replay, BGAD-Replay) and staged approaches (DRAEM-CLS, BGAD-CLS).