Text-Guided Image Segmentation

Text-guided image segmentation studies how to turn class names, referring expressions, clinical reports, and instructions into precise pixel-level masks. Common approaches align language and visual features with contrastive learning or cross-attention, condition segmentation models such as Segment Anything with text prompts or lightweight adapters, and use generative models or multimodal reasoning to propose and refine masks. These methods support settings from open-vocabulary recognition to medical and remote-sensing segmentation, with training-free, few-shot, and supervised approaches. A central challenge is grounding varied or ambiguous language in accurate boundaries across domains while balancing robustness, reliability, and computational cost.