Local self-similarity based backprojection for image upscaling

Abstract

This paper proposes a new local self-similarity based backprojection (LSSBP) approach for the image upscaling problem. A backprojection estimates a high resolution image by backprojecting reconstruction error from a low resolution image. The reconstruction error, calculated in the low resolution domain, needs to be anisotropically upscaled in order to reduce artifacts. By using a self-similarity between low and high resolution images, a high frequency data of the high resolution image can be transferred from its similar counterparts in the low resolution domain without many artifacts. However, the direct combination of the backprojection and the self-similarity based upscaling suffers from high computational complexity because similar patches can exist in the entire image and across scales. In this paper, in order to localize search region of similar patch, the local self-similarity based upscaling is adopted for the backprojection of reconstruction error. Therefore, in the backprojection framework, the generated high resolution image has consistency with the low resolution image, and more clean and sharp edge can be obtained by using the local self-similarity. The experimental results show advantages of the proposed approach. The proposed approach can be applied to resolution improvement of surveillance and intelligent vehicle vision systems where images are captured by low resolution sensors.