Few-Shot Fabric Defect Detection via Vision-Text Paired In-Context Learning and Two-Stage Reflection

Abstract

Industrial visual sensing systems play a critical role in automated quality inspection. However, deploying data-driven perception models on industrial visual sensors faces significant bottlenecks due to the extreme scarcity of annotated anomaly samples and the pronounced long-tailed distribution of defect types. These factors render traditional supervised models prone to overfitting and catastrophic failure when sensing rare anomalies. To address these limitations in the sensor data processing pipeline, this article proposes a novel few-shot perception framework leveraging multimodal large language models (MLLMs) as a cognitive backend, without task-specific fine-tuning of the sensing model. Our approach introduces two key innovations: 1) a vision-text paired in-context learning (VTP-ICL) mechanism that constructs structured multimodal prompts to activate pretrained knowledge for precise defect localization and 2) a two-stage reflection pipeline that decouples initial prediction from verification, enabling the model to self-correct outputs based on explicit criteria. Extensive experiments on a real-world industrial fabric dataset demonstrate that our method achieves state-of-the-art performance under strict few-shot conditions (two shots per class). Notably, our approach maintains an exceptional balance between precision and recall, significantly outperforming traditional detectors, particularly on hard-to-sense, low-contrast defects. These results validate the potential of MLLMs as robust and generalizable engines for next-generation industrial sensing applications.