A CLIP-driven referring image segmentation model for Mini LED defect inspection

Mini light-emitting diode (Mini LED) technology has been extensively adopted in next-generation displays, owing to its superior luminance, high energy efficiency, and precise pixel-level control. However, various defects arise during Mini LED manufacturing, seriously compromising product quality. Thus, reliable detection approaches are urgently needed. Yet, most existing unimodal vision-based methods only produce fixed-form outputs, which limits flexibility and interactive analysis. To mitigate this, we introduce referring image segmentation (RIS), a language-guided semantic segmentation method, into Mini LED defect detection. We construct a large-scale dataset with 55,809 image-language-mask triplets and expand the output task space from a single type to 13 formats. We further propose MLED-CUNet by integrating CLIP with U-Net, enabling prompt-driven segmentation of Mini LED images. Unlike conventional RIS approaches that mainly output binary masks, our method supports multi-class segmentation across 18 categories, covering a broader set of inspection tasks. Experiments show that the proposed approach achieves 60.26% AmIoU (Average mean Intersection over Union) across 13 tasks, where AmIoU represents the average of the mean Intersection over Union (mIoU) values for each task. This supports clearer visualization of Mini LED components and accurate detection of defects with diverse attributes.