Research On Denosing And Super-resolution Of Hand Depth Images

Hand depth images are depth images with the three-dimensional information of human hands. Compared to the traditional gray and color images of hands, hand depth image are more robust to noise of environment. However, due to the environmental lights, occlusion and other factors, depth images provided by traditional devices such as Kinect or To F cameras have lower resolutions and contain much noise and singular pixels. Improving the quality of hand depth images is important for their applications in the fields of computer vision, computer graphics and human-machine interaction.Traditional approaches of depth image denoising and super-resolution can be roughly divided into three categories: filtering approaches, reconstruction model based approaches and sparse representation based approaches. In general, filtering approaches impose adaptive smoothing filters on depth images based on neighborhood of pixels, and do not treat well at the silhouettes of hands due to the selection of neighborhoods of pixels. Reconstruction model based approaches build probabilistic models or assembling models based on a large number of training data. Such approaches are more suitable for treating depth video instead of a single depth image. Sparse representation based approaches essentially obtain a group of overcomplete bases(called dictionary) of the original image based on training data, and produce the original image or higher-resolution image using the sparse combination of the dictionary. Within these approaches, the procedure of dictionary training is key to the quality of image denoising and super-resolution. Because the hand depth images suffer from low-resolution and high noise, traditional sparse representation based approaches fail in denoising and super-resolution. We propose a two-stage depth image denoising and super-resolution which combines a RGB-D based bilateral filters and robust dictionary training. The main contribution of this paper includes the following two aspects1. This paper proposes a joint bilateral filter using both color image and depth image to denoise depth image. It removes most of the noise in depth image based on the relevance on the color image and continuity on the depth image with a more strong constraint. As some noise still exists, we propose image denoising via sparse representation using a robust dictionary obtained by an improved orthogonal matching pursuit. Experimental results show that this method can effectively improve the efficiency.2. This paper proposes image super-resolution using sparse representation over coupled dictionary learning to hand depth images. We obtain a high resolution dictionary and a low resolution dictionary using the KSVD method, whosedictionaries are obtained from the improved orthogonal matching pursuit. Then we compute the sparse coefficient of low resolution depth image and low resolution dictionary. We obtain high resolution depth image through high resolution dictionary multiply the sparse coefficient. Experimental results show that our method can obviously improve objective qualities of super-resolution of depth images.