Mask-Guided Super-Resolution and Diffusion-Based Restoration for Low-Quality Depth Images

Under the constraints of limited acquisition device capabilities, depth images are often affected by noise contamination and loss of human-related details, which compromises their applicability in computer vision tasks. This paper proposes a restoration method for low-quality depth images based on diffusion models and super-resolution techniques. We employ a pre-trained human detection model combined with manual assistance to segment human masks in both clean and noise-corrupted regions. In the image enhancement module, we analyze depth values to identify distant regions and leverage the human mask as a constraint to guide the super-resolution reconstruction process, thereby recovering global image details. In the image restoration module, the super-resolved image serves as global prior information, while the segmented human mask provides local structural cues to guide the diffusion model in repairing degraded depth images, aiming to improve overall image quality and structural consistency. Experimental results demonstrate that the proposed method significantly improves human detection performance on the restored depth images. Compared to the original images, introducing the human mask increases the mAP from 0.427 to 0.689, Precision from 0.61 to 0.93, and Recall from 0.75 to 0.86, yielding improvements of 0.262, 0.32, and 0.11, respectively. These results validate the effectiveness of the mask-guided mechanism in the context of image restoration.