Visual Storage Battery Inspection via Large Vision-language Model

Storage batteries are widely used across various industries, and their failure can lead to severe consequences. With the advancement of artificial intelligence, particularly computer vision, visual data of storage batteries can be captured through monitoring cameras and processed using vision-centric convolutional neural networks (CNNs) or transformers for storage battery inspection. However, traditional vision-centric methods often rely on large amounts of annotated data, which is a significant limitation in the field of storage battery inspection. Additionally, storage batteries typically operate in complex environments with severe noise, leading to lower recognition accuracy in conventional approaches. To address these challenges, a specialized dataset tailored for storage battery inspection is created. We then introduce VLMSBI, a novel vision-language model-based method, to perform storage battery inspection. Our approach demonstrates superior performance compared to traditional vision-centric CNN- and transformer-based methods, while utilizing only minimal annotated data for supervised fine-tuning (SFT). Notably, VLMSBI's versatile architecture allows for seamless adaptation to other industrial component inspection tasks by simply modifying the fine-tuning data and corresponding input prompts.