Resource-Efficient Adaptation of Large Vision–Language Models for Multimodal Defect Inspection in LNG Tank Manufacturing

Automated inspection in LNG tank manufacturing remains challenging due to the need for simultaneously identifying defect regions and generating structured, con-text-aware reports—capabilities not provided by existing detection models. This study presents the first systematic investigation into whether recent large vision–language models (LVLMs) can perform unified defect localization and metadata prediction within this industrial domain. Using a carefully balanced 15,000-image subset of the AIHub LNG Tank Quality Inspection dataset, seven state-of-the-art LVLMs from the Qwen, Gemma-3, and LLaVA families were fine-tuned via QLoRA to operate under a 24 GB GPU constraint. Performance was comprehensively evaluated using mAP@50, F1-scores for seven semantic fields, and BLEU for JSON-structured report fidelity. Qwen3-VL-8B achieved the strongest overall results, delivering high-precision localization (mAP@50 = 87.42%) and consistent multimodal report generation. Qualitative analyses further highlight the model’s strengths in defect recognition while identifying challenging categories such as cracks and scratches. Additional comparisons with YOLO and RT-DETR detectors demonstrate that LVLMs can approach the accuracy of dedicated detectors while providing richer and more interpretable outputs. The findings introduce LVLMs as a novel and promising direction for industrial inspection systems requiring both spatial precision and structured semantic reasoning.