Super-resolution of text images through neighbor embedding

Single-image super-resolution (SISR) is the problem ofgenerating a high resolution (HR) image from a single low resolution (LR) image, possibly with the help of a set of training images. The SISR technique known as neighbor embedding (NE) is based on the assumption that corresponding small patches in low and high resolution versions of an image form manifolds with similar local geometry. NE utilizes a training ensemble of pairs of low and high resolution image patches, where the patches in a given pair represent the same image region. An input patch from a LR image is approximated by nearby LR training patches and a HR patch estimate is constructed from the corresponding combination of HR training patches. While NE has shown good success for super-resolution of face and scene imagery, little has been reported on NE for enhancing text images. We apply NE to enhance LR text images, and achieve good HR estimates at 2x, 3x and 4x magnification. Our experiments show that NE raises PSNR found with bicubic interpolation (BI) by 68% and 89% at 3x, and 4x magnification, respectively. We show how to naturally extend the original NE luminance features to an arbitrary number, and achieve further improvement in PSNR by adding just one more feature.