Depth-Map-Assisted Texture and Depth Map Super-Resolution

摘要

With the development of video technology, high definition video and 3D video applications are becoming increasingly accessible to customers. The interactive and vivid 3D video experience of realistic scenes relies greatly on the amount and quality of the texture and depth map data. However, due to the limitations of video capturing hardware and transmission bandwidth, transmitted video has to be compressed which degrades, in general, the received video quality. This means that it is hard to meet the users’ requirements of high definition and visual experience; it also limits development of future applications. Therefore, image/video super-resolution techniques have been proposed to address this issue. Image super-resolution aims to reconstruct a high resolution image from single or multiple low resolution images captured of the same scene under different conditions. Based on the image type that needs to be super-resolved, image super-resolution includes texture and depth image super-resolutions. If classified based on the implementation methods, there are three main categories: interpolation-based, reconstruction-based and learning-based super-resolution algorithms. This thesis focuses on exploiting depth data in interpolation-based super-resolution algorithms for texture video and depth maps. Two novel texture and one depth super-resolution algorithms are proposed as the main contributions of this thesis. The first texture super-resolution algorithm is carried out in the Mixed Resolution (MR) multiview video system where at least one of the views is captured at Low Resolution (LR), while the others are captured at Full Resolution (FR). In order to reduce visual uncomfortableness and adapt MR video format for free-viewpoint television, the low resolution views are super-resolved to the target full resolution by the proposed virtual view assisted super resolution algorithm. The inter-view similarity is used to determine whether to fill the missing pixels in the super-resolved frame by virtual view pixels or by spatial interpolated pixels. The decision mechanism is steered by the texture characteristics of the neighbors of each missing pixel. Thus, the proposed method can recover the details in regions with edges while maintaining good quality at smooth areas by properly exploiting the high quality virtual view pixels and the directional correlation of pixels. The second texture super-resolution algorithm is based on the Multiview Video plus Depth (MVD) system, which consists of textures and the associated per-pixel depth data. In order to further reduce the transmitted data and the quality degradation of received video, a systematical framework to downsample the original MVD data and later on to super-resolved the LR views is proposed. At the encoder side, the rows of the two adjacent views are downsampled following an interlacing and complementary fashion, whereas, at the decoder side, the discarded pixels are recovered by fusing the virtual view pixels with the directional interpolated pixels from the complementary downsampled views. Consequently, with the assistance of virtual views, the proposed approach can effectively achieve these two goals. From previous two works, we can observe that depth data has big potential to be used in 3D video enhancement. However, due to the low spatial resolution of Time-of-Flight (ToF) depth camera generated depth images, their applications have been limited. Hence, in the last contribution of this thesis, a planar-surface-based depth map super-resolution approach is presented, which interpolates depth images by exploiting the equation of each detected planar surface. Both quantitative and qualitative experimental results demonstrate the effectiveness and robustness of the proposed approach over benchmark methods.