Image Super-resolution Based On Clustering Oriented Convolutional Neural Networks

Human vision has played a significant role for human beings in obtaining information,which motivates the development of machine vision and modern information technology.With the upgrading of image display devices,low-quality images can hardly meet the demand for high-quality visual effect.The variation of performance of image acquisition equipment leads to images in a variety of resolution.Thus,image super-resolution came into being and had become a classical task in machine vision.The goal of image super-resolution is to generate images in high-resolution from images in low-resolution based on the learning priors.According to the priors,existing image superresolution algorithms can be categorized into four classes: prediction models,edge based methods,image statistical methods and patch based methods.Both of these methods usually involve sparse representation,support vector regression and other classical machine learning algorithms,while the emergence of deep learning based supervision approaches almost subvert all of them.As one of the deep learning based image super-resolution family,the convolutional neural network based approach regards the low-resolution images and the corresponding highresolution images as inputs and labels,respectively,to train the convolutional neural network.In that case,images in low-resolution can be reconstructed to high-quality images by utilizing the convolutional neural network based super-resolution models.We propose two image superresolution methods incorporating the convolutional neural network based frameworks: image super-resolution based on spatial domain clustering oriented convolutional neural networks and image super-resolution based on frequency domain clustering oriented convolutional neural networks.Image super-resolution based on spatial domain clustering oriented convolutional neural networks crop the high-resolution images in training dataset into image patches and cluster them using K-means algorithm,then image patches in each cluster are utilized to train their corresponding convolutional neural network models.In the stage of super-resolution,each lowresolution input image is cropped into patches and each patch is assigned to the related cluster based on spatial similarity.Its high-resolution image patch is reconstructed by using the corresponding convolutional neural network model.All high-resolution patches are then integrated into one high-resolution image.Image super-resolution based on frequency domain clustering oriented convolutional neural networks extract different frequency components of high-resolution images using wavelet transform.Low-resolution images and the frequency components extracted from their related high-resolution images are treated as inputs and labels to train convolutional neural network models.Those models are employed to generate the frequency components of a highresolution image from a given low-resolution image.Then the frequency components are adopted to reconstruct the high-resolution image by inverse wavelet transform.Image super-resolution based on spatial domain clustering oriented convolutional neural networks solves the problem of “one-to-many” and achieves the non-linear mapping between low-resolution images and high-resolution images.Image super-resolution based on frequency domain clustering oriented convolutional neural networks proposes a novel algorithm to enhance image resolution based on different frequency components.Those two algorithms presented in this paper provide new insights of image super-resolution,both of which have high application value and potential for further research.