Super-resolution Image Reconstruction by High-frequency Components Estimation Based on Self-reference

Abstract

Although high-resolution (HR) image acquisition devices have been developed, such devices cannot always provide high-quality or HR images because of environmental limitations, such as power consumption and bandwidth limitation. Therefore, several image resolution enhancement techniques have been developed to advance the performance of digital display devices recently. In this thesis, a novel technique for reconstructing an HR image from a single low-resolution (LR) image is proposed. The proposed algorithm consists of two main parts: high-frequency component estimation that uses self-reference and a discrete cosine transform (DCT)-based up-scaling scheme. In the first step, we estimate the high-frequency components using self-reference from the input LR image size conversion. In the second step, a DCT-based up-scaling scheme is utilized to reconstruct the HR image. The advantages of the proposed method are summarized as follows: First, the proposed method only requires one LR input image. Second, the proposed method is non-parametric, which means that it does not require any parameters, such as dictionary information (e.g., LR-HR pair information for learning-based super-resolution approaches) or priors (e.g., gradient profile priors for edge-based super-resolution approaches). Third, the proposed method can be applied directly to image and video compression systems, because it is based on DCT. Fourth, the proposed method is a real-time super-resolution algorithm, which means that it can be used in various applications in the real world. In order to evaluate the performance of the proposed method, we use the peak signal-to-noise ratio (PSNR) and the structural similarity (SSIM) as the quantitative measures. Experimental results show that the proposed method achieves better qualitative/quantitative performance compared with existing methods, especially in terms of texture and detail.