VLCIM: A Vision-Language Cyclic Interaction Model for Industrial Defect Detection

Abstract

Accurate defect detection is an important element in ensuring product quality and safe equipment operation. However, due to the lack of deep cross-modal interactions during vision feature extraction, existing methods often suffer from attention bias, which ultimately limits detection accuracy. To address this issue, this paper proposes a Vision-Language Cyclic Interaction Model (VLCIM), which progressively optimizes vision feature extraction by integrating domain prior knowledge and generic large model, effectively bridging the dual-domain barrier between “generic-specific” and “vision-language”. Specifically, progressive cyclic interaction learning is proposed for the first time, which integrates a recursive guidance module (RGM) and cross-modal interaction (CMI) strategy to realize bidirectional dynamic fusion and collaborative optimization of vision and language features. Furthermore, the proposed dual-view synergistic detection mechanism enhances discriminative decision responses, significantly improving the model’s boundary perception ability and decision-making accuracy in complex scenarios. VLCIM achieves high-precision defect detection by establishing a cyclic interaction mechanism between domain-specific language features and vision representations. Experimental results on three industrial datasets demonstrate that VLCIM achieves improvements of 5.9%, 5.6%, and 4.1% in mIoU over the state-of-the-art (SOTA) methods, indicating its validity and generalization in different scenarios.