SAPOCLIP: A Self-Adaptive Prompt Optimization Framework for Industrial Anomaly Detection

Industrial anomaly detection (IAD) is crucial for enhancing production efficiency and quality. Although large vision-language model based IAD methods, such as CLIP-based IAD methods, performed excellently in zero/few-shot performance compared to fully supervised ones and showed great potential to improve IAD, the exploration of text information processing was rarely available. In this paper, we focused on the text information processing aspect of CLIP and proposed a self-adaptive text prompt optimization framework, SAPOCLIP, which was implemented using State Description Parameter (SDP) and Main Prompt Template (MPT). Comprehensive experiments illustrated that the performance of the SAPOCLIP in few-shot anomaly classification and segmentation tasks was significantly improved. For image-level/pixel-level anomaly detection tasks, its performance under 1/2/4 few-shot settings surpassed that of other SOTA few-shot task models remarkably.