New AI framework combines language models with dynamic expert networks to improve visual defect detection while reducing computational costs

DS-MoE framework integrates distilled large language models with sparse Mixture-of-Experts architecture for adaptive expert activation based on semantic relevance

Text-guided dynamic routing aligns textual semantics with visual patterns to resolve inter-class ambiguity and improve recognition accuracy

Lightweight MobileSAM encoder enables real-time inference while preserving multi-scale detail for defect analysis

Framework tested on PCB, aluminum foil, and mold defect detection datasets, addressing challenges of high inter-class similarity and extreme scale variation

Addresses limitations of existing rigid fusion mechanisms and heavy annotation pipelines to achieve better generalization across diverse real-world data