Single Image Super Resolution Based On Edge-Prior

Single image super resolution aims to estimate high resolution image from the low resolution one, in order to reconstruct the high frequency and details when down-sampled. Approaches can be grouped into four categories: interpolation-based, learning-based, reconstruction-based, and edgedirected methods.In the thesis the edge-directed super resolution methods are mainly studied, and two methods of edge-directed are presented. The two contributions and work are summarized as follows:1. The edge directed image super resolution through the linear regression of examples is presented, including two gradient estimation methods: gradient estimation based on ridge regression and gradient estimation based on linear mapping function. In gradient estimation based on ridge regression, recognizing that the training samples of the given sub-set for regression should have similar local geometric structure based on clustering, this section first employs high frequency of LR image patches with removing the mean value to perform such clustering. In gradient estimation, the coefficients of patch sparse filtering feature to the high resolution sample features of the same cluster are calculated through the ridge regression. And HR gradient of the patch is estimated with the coefficients and HR gradient of samples. Moreover, the pre-computed projective matrix of the ridge regression can reduce the computational complexity further. In gradient estimation based on linear mapping function, the mapping function is learned to predict high resolution gradient of patches from gradient of bicubic patches. For bicubic patches, the high resolution gradient can be calculated by the mapping function prior.2. Image super resolution on multi-domain via regression based on extreme learning machine is presented, which includes two parts: the estimation of gradient domain and high frequent components, the reconstruction model of multi-domain. In training part, for clustered samples, high frequency components from high resolution images as the output values and the patches with removing the mean as the input are first fed into ELM to learn models. Then the estimating error of high frequency component, the gradient of horizon and vertical direction, the gradient of diagonal direction as the output values, the patches with removing the mean and the frequency component estimation are jointed as the input, are fed into ELM to learn models. In reconstruction, the gradient of horizon and vertical direction, the gradient of diagonal direction are set as the gradient constraint in gradient constraint, in order to obtain the initial reconstructed image with sharp edges; Then the estimated high frequency image is set as the first constraint on image domain, and the initial reconstructed image is set as the second constraint. At the same time, the high frequency of reconstructed image after one iteration is processed by non-local means, in which self-similarity is used. The experiments show the joint of gradient domain and high frequency domain can make the reconstructed images with sharper edges and fine details.