Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

IAD-R1: Reinforcing Consistent Reasoning in Industrial Anomaly Detection

Created by
  • Haebom

Author

Yanhui Li, Yunkang Cao, Chengliang Liu, Yuan Xiong, Xinghui Dong, Chao Huang

Outline

This paper proposes IAD-R1, a novel post-training framework that leverages the Vision-Language Model (VLM) to address the problem of anomaly detection in industrial settings. To address the lack of defect data, we employ a two-stage training strategy. The first stage, Perception Activation Supervised Fine-Tuning (PA-SFT), utilizes the high-quality Chain-of-Thought dataset Expert-AD to enhance anomaly detection and establish inference-answer correlations. The second stage, Structured Control Group Relative Policy Optimization (SC-GRPO), further enhances anomaly detection through a reward function. Experimental results demonstrate that IAD-R1 improves performance on seven VLMs, particularly on the DAGM dataset, achieving an average accuracy improvement of 43.3% over the baseline model. Furthermore, a 0.5B parameter model trained with IAD-R1 outperforms commercial models such as GPT-4.1 and Claude-Sonnet-4 in zero-shot settings. The code, dataset, and model weights are publicly available.

Takeaways, Limitations

•
Takeaways:
◦
We present a novel post-training framework, IAD-R1, that significantly improves VLM-based industrial anomaly detection performance.
◦
Versatility applicable to various VLM architectures and parameter sizes
◦
Achieving performance that surpasses commercial models in zero-shot settings
◦
Demonstrating the Effectiveness of Expert-AD, a High-Quality Chain-of-Thought Dataset
◦
Increase research reproducibility and scalability by making code, datasets, and model weights public.
•
Limitations:
◦
There is a possibility that the performance improvement of IAD-R1 may be biased towards a specific dataset (DAGM).
◦
Need to verify generalization performance for other industries or types of abnormalities
◦
Possible lack of detailed description of the creation process and quality of the Expert-AD dataset
◦
Additional explanation is needed regarding the design of the reward function of SC-GRPO.
👍
Made with Slashpage