A spatial inter-view auto-regressive super-resolution scheme for multi-view image via scene matching algorithm

Abstract

Binocular suppression theory states that the stereo vision quality is not much influenced by asymmetric degradation of the individual views. Based on these findings, mixed resolution (MR) multi-view framework jointly utilizes the lower and full resolution images to reduce the data amount, while maintaining good stereo vision quality. To enhance the resolution of the lower resolution image, a novel superresolution scheme for the MR multi-view framework is presented in the paper. It is based on the assumption that the image is modeled as a 2D piecewise auto-regressive process. In the scheme, each pixel to be interpolated is estimated as the linear weighted summation of the pixels, which are consisted of the spatial neighboring ones from lower resolution image in the current view and the ones from the full resolution image in the neighboring views. To get the corresponding pixels in the neighboring views that match the scene to be reconstructed, a window based scene-matching approach is used. Through exploiting the spatial correlation and the inter-view correlation, the proposed scheme achieves a significant gain in PSNR and visual quality for the test sequences.