Abstract
Nowadays, High-resolution (HR) monitors have been very popular in the market. However, many legacy images/videos are stored in low-resolution (LR) formats These LR files need to be enlarged before playing on HR monitors. Super-Resolution(SR) algorithms have been proposed to enhance the image resolution. The most common way to do up-scaling is using interpolation-based methods. These kind of methods are simple but may suffer from noticeable blurring edges or unclear textures. Human vision is more sensitive to these high-frequency(HF)-rich regions, so these regions should not be processed only by interpolated method. Example-based method is one of the most popular approaches, which aim to enhance the HF of interpolated results. The concept of these methods are to make use of the correspondence between LR and HR image. In Example-based methods, the LR patches are replaced by their HR versions. These HR patches are selected by similar patch search process. Hence, whether suitable HR patch can be found in the patch search have a significant impact on the final result. The effect of search methods depends on what they use in comparison. For example, the common way to search for similar patches is comparing the low-frequency(LF) values. However, edge and texture regions are not suitable since only raw pixel information is used. The radient-based patch search is proposed to enhance the effect on these HF regions. Nevertheless, it’s hard to compare the gradient difference on smooth regions. In this paper, We proposed a hybrid patch search process that combine the gradient and LF-based patch search for further enhance the effect of above methods. We use the assumption of local self-similarity to limit the search area within a small window but get similar results in most cases. In the proposed framework, two different patch search methods are applied. For edge regions, we use the gradient-based patch search; in smooth regions, low frequency-based patch search is adopted. When the difference is close between two patches of the hybrid patch search, we further compare the gradient direction for verification. In the experimental results, comparing with the SR methods which use LF-based patch search only, our proposed method get 9.5% higher at average PSNR values and 3% higher at SSIM average values. The proposed method gets 10% higher at average PSNR values and 5% higher at SSIM average values compared with the SR methods which use gradient-based patch search only. At the same time, we simplify the patch verification process. The computation for HF reconstruction is reduced about 54% compared to the gradient-based SR method.