Research On Super-resolution Image Restoration Based On Learning And Total Variation Regularization

The resolution of the images represents the richness of image details. It is animportant subject to improve the resolution of the image in the field of imageprocessing and computer vision. Super-resolution image reconstruction (SISR) is aone of the methods to improve the image resolution. It breaks through the traditionalmethod which relies on the improvement of the performance of the hardwareequipment. It reduces the cost of imaging procedure by relying on softwaretechnology. At present, this method has been shown the huge development potential inthe remote sensing, medical imaging, video surveillance, digital media and manyother areas. It is a very meaningful to research new super-resolution restorationtechniques.The Total variation has been widely used in image processing due to its ability topreserve image edges in image processing. The learning-based super-resolution imagerestoration method is an important research direction for the super-resolution imagerestoration. This paper mainly studies the learning-based super-resolutionreconstruction with minimizing total variation for single image super-resolutionreconstruction problem. The main content of the paper and conclusions are as follows:1. In this paper, a novel super-resolution image restoration model based onlearning and total variation regularization is proposed. Total variation (TV) isintroduced to the learning-based image super-resolution reconstruction method.2. We design an alternating iteration algorithm to solve the proposed model. Byintroducing new variables, the model is reformulated to two optimization problemswhich are easier to be solved. Meanwhile, the feature of the image is considered inthis paper. Nonlocal self-similarity of images and iterative back projection algorithmare used to process the recovered image.3. Experimental results on different images demonstrate the effectiveness of theproposed model for the super-resolution image restoration, the quality of imageobtained by our model is better than the one by the traditional interpolation algorithmand the method based on sparse coding based super-resolution method both in thePSNR value and the visual effect.