CHIPS: Human-AI collaborative prompt generation for industrial defect segmentation

Industrial quality control demands accurate defect segmentation with minimal human effort, yet current methods sacrifice either automation efficiency or segmentation precision. We propose CHIPS (Collaborative Human-AI Intelligent Prompt Segmentation), a framework that bridges few-shot learning and the Segment Anything Model (SAM) to achieve both automation and accuracy. CHIPS automatically transforms coarse few-shot predictions into precise segmentations through intelligent prompt generation, mimicking expert interaction patterns without manual intervention. The system operates in dual modes: fully automatic mode for high-throughput inspection, and collaborative mode that routes challenging cases to human experts through programmatic interfaces. Our uncertainty estimation mechanism identifies ambiguous cases, enabling flexible deployment strategies based on quality requirements. Evaluated on CID-20, a comprehensive 20,344-image industrial defect dataset spanning 20 product categories, CHIPS achieves 45.93% mIoU in 7-shot settings—a 1.88% improvement over state-of-the-art baseline. The framework demonstrates consistent gains averaging 2.06% across all shot configurations while achieving 3.7–8.1 FPS processing speed. By seamlessly integrating intelligent automation with selective human expertise, CHIPS establishes a new paradigm where industrial quality control achieves both operational efficiency and segmentation precision, adapting to diverse manufacturing requirements from fully automated to quality-critical human-verified workflows.