An effective spatial-temporal denoising approach for depth images

Image retrieval and computer vision rely heavily on noise removal and image hole padding to ensure the accurate rendering of images from depth cameras. Noise can undermine the accuracy of depth values, resulting in object occlusion and temporal variation in depth pixels. This paper proposes an efficacious approach to dealing with this problem based on an exemplar-based inpainting method for the removal of noise from object-removed images obtained using RGB-D cameras. The proposed method was also applied to the Tsukuba Stereo database, which provides 3D video with ground truth disparity maps. Experiment results were evaluated in terms of peak signal-to-noise ratio (PSNR) and computation time. Moreover, the self-recorded RGB-D depth images database was used to verify the corrections in spatial and temporal variation of depth values. Our results demonstrate the effectiveness of the proposed method by comparing the original occluded depth images with the images being processed.