SurgiMist: Controlled Smoke Synthesis for Surgical Image Restoration and Perception

Surgical smoke changes image appearance and obscures cues used by both restoration models and downstream perception systems. Paired clear and smoky observations are difficult to acquire in surgery, while restoration scores alone do not establish whether removing smoke helps a task. We introduce SurgiMist, a controlled synthesis framework that combines relative depth, heterogeneous optical fields, layered accumulation, texture evolution, and scattering-inspired image formation. Its nested configurations support restoration benchmarking, and a geometry-preserving variant transfers existing semantic labels to smoke-augmented images. We evaluate the framework through a fixed restoration benchmark, three restoration backbones on external data, and a video-disjoint surgical segmentation study with three training seeds. On the fixed four-condition restoration benchmark, the complete configuration achieves the best mean PSNR, SSIM, and LPIPS among the four synthesis configurations. For segmentation, its geometry-preserving variant reaches 50.18 ± 0.12% mean intersection-over-union across 15 balanced synthetic conditions, a 15.71 percentage-point gain over original-image training. Under severe smoke, direct augmentation reaches 50.28%, compared with 39.45% for a learned restoration front end. In a blinded review of 500 paired judgments by five doctors on real-smoke frames, B4 receives slightly more votes than LSD3K for visibility, structure preservation, and viewing preference, and fewer artifact flags. Matched component controls separate the benefit of smoke augmentation from Gaussian smoothing alone. Together, these results position controlled synthesis as a practical interface for measuring restoration quality and training smoke-robust perception, with the downstream objective determining the useful processing route.