PROJECT TITLE :

Deep Color Guided Coarse-to-Fine Convolutional Network Cascade for Depth Image Super-Resolution

ABSTRACT:

The task of super-resolution of depth images is both significant and difficult. In order to deal with this issue, we present a new deep color-guided coarse-to-fine CNN framework in this research. There is a data-driven way to approximate the optimal depth image super-resolution filter, rather than hand-designed filters, which we present first. For upsampling depth images, the filter learned from Big Data samples is more precise and reliable. Second, we introduce a coarse-to-fine CNN to learn filter kernels of varying size. For a crude high-resolution depth image, the CNN learns larger filter kernels at the coarse stage. The coarse high-resolution depth image is used as the input for the fine stage so that smaller filter kernels can be trained in order to achieve more precise results. Our ability to progressively recover high frequency details thanks to this network is much enhanced. A colour guiding technique that incorporates colour difference and spatial distance for depth picture upsampling is then developed. A high-resolution colour map is used to modify the interpolated high-resolution depth image. In order to avoid texture copying artefacts and maintain edge features, the depth of high-resolution image obtained can be guided by colour information. Our depth map super-resolution performance has been demonstrated in quantitative and qualitative experiments.