Contour-aware diffusion model for restoring atmospheric turbulence-degraded facial images

Atmospheric turbulence causes spatially non-uniform blur and geometric distortion in long-range facial imaging, posing substantial challenges to reliable image restoration. To address these challenges, this paper proposes a diffusion model incorporating a contour prior and channel attention to restore facial images degraded by atmospheric turbulence. We construct a prior extraction network, VGG-PriorFace, to extract contour information from turbulence-degraded facial images, assisting the diffusion model in establishing a more realistic image distribution. Introducing a channel attention mechanism into the diffusion model, enables the framework to preferentially utilize high-dimensional features from intermediate layers. Experimental results show that our improved scheme significantly boosts the visual perception of the restored images and enhances the clarity of facial features, contours and the background.