Multi-domain transformer IQA algorithm for AOI optical scheme evaluation

Le résumé fourni par la source

Automatic Optical Inspection (AOI) systems require carefully designed optical configurations to ensure adequate defect visibility during image acquisition. In industrial practice, engineers typically compare images captured under different optical settings and manually select the configuration that best exposes defects. However, this process is inherently subjective and lacks quantitative modeling, making systematic optimization of optical setups challenging. In this work, we formulate AOI optical scheme selection as a supervised learning problem by modeling engineer-annotated defect exposure scores. To support this study, we construct the WAFER-OS100 dataset, which contains wafer images acquired under 100 distinct optical configurations. To effectively capture subtle defect characteristics in complex industrial environments, we propose a task-oriented learning framework that integrates frequency-aware image representations with explicit optical parameter encoding within a transformer-based architecture. By jointly modeling image content and optical parameters, the proposed method learns the relationship between optical configurations and defect exposure quality, enabling data-driven evaluation and ranking of imaging schemes. Experimental results demonstrate a strong correlation between the predicted scores and engineer annotations, indicating that the proposed framework can effectively approximate human evaluation and provide practical support for AOI optical scheme evaluation and ranking.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.