Abstract
A method of super-resolution of a single image is presented for approximately minimizing image reconstruction errors. A high-resolution image is estimated with an approximate Newton method for minimizing the error between an input low-resolution image and smoothed high-resolution image with a reduced size. For smoothing the high-resolution image, we use an anisotropic bilateral filter with a spatial weight of DOG (Difference Of Gaussians) characteristics and additionally we adopt the unsharp masking for enhancing the edge-preservation property of the filter. We examine with experiments the effectiveness of spatial anisotropy of the bilateral filter and utilization of the mode filter and compare the PSNR and SSIM of reconstruction errors with previous methods.
