Accurate and Fast Degradation Function Estimation for Single Image Super-Resolution

Abstract

Single image super-resolution is a long-lasting ill-posed problem which seeks attention from both academia and industry. The up-scaling the low-resolution images by digging out high-frequency details is still yet to be studied deeply. In the scope of this thesis we proposed a full pipeline for single image super-resolution. Our research question was to effectively estimate the degradation function of the image while capturing. The degradation function is then used to super-resolve the images with sharp edges and without artefacts. A close estimate for degradation function is indispensable for super-resolving images. \\

In this thesis we propose an effective and fast way of estimating the degradation function. Our novelty lies in the degradation function estimation part where we empirically proved that there is a significant relationship between the degradation function and the estimated blur kernel on the low-resolution image. Rooted to this finding, we first estimate the degradation function and then feed this estimated degradation function alongside with the low-resolution image to a non-blind super-resolution module. The whole pipeline has the modular and fast property for up-scaling the low-resolution images while preserving and promoting high-frequency components hidden in the low-resolution sample. Extensive experiments illustrated that our degradation function estimation algorithm is faster and more accurate compared to the existing algorithms in the literature. Connecting the degradation function estimation module to a non-blind super-resolution module, we supersede the state-of-the-art super-resolution algorithms in the literature. The future work involves enhancing the robustness of the degradation function estimation module to noisy inputs and having a fully-connected end-to-end training which includes the fine-tuning of the non-blind super resolution modules.