Image Super Resolution Based On Reconstruction Method

In this age of change, the human society is progressing rapidly on various fronts. Simple images cannot satisfy all people’s requirements. Digital image processing technology plays an significant role in meeting with different people’s requirements in processing image. In many digital image application fields, high resolution images are often desired. High resolution means that pixel density in an image is high, thus a high resolution image can provide more image details that may be crucial in many image applications. For this reason, finding a way to increase the image resolution level is of great importance.Image super resolution is a process that estimates a high resolution image from one or multiple low resolution images. In comparison with the situation when multiple images are input, single image super resolution is more complicated due to the unknown blur kernel and lacking of enough low resolution images.There are three main ways to deal with image super resolution problem: interpolation-based methods, learning-based methods and reconstruction-based methods. Interpolation-based methods need to construct an interpolation function, and then apply the interpolation function to estimate a high resolution image from a given low resolution image. Learning-based methods realize the image super resolution process by learning a relationship between low resolution images data sets and high resolution images data sets. Strictly speaking, due to the introduction of additional training data sets, learning-based methods are not single image super resolution approaches. Through studying the formation process that how a low resolution image is produced from a high resolution, reconstruction-based methods establish the model to achieve the image super resolution process. In this paper, we mainly focus on reconstruction-based methods.Inspired by a fast image upsampling method, we study the situation that the blur kernel is unknown. In this paper, we introduce a fast single image super-resolution method based on deconvolution strategy and kernel estimation. In the deconvolution process, we take a fast total variation deconvolution method to improve the running time of the algorithm. Besides, because the blur kernel is unknown in image super resolution process, we utilize an iterative strategy to correct the kernel and narrow the gap between the initial kernel and the real one. In the numerical experiments, we compare the proposed approach with current state-of-the-art approaches. The experimental results demonstrate that our method can improve the visual effect of the high resolution image. In addition, the running time of the proposed method is very fast.