Learning-Based Nonparametric Image Super-Resolution

We present a novel learning-based framework for zooming and recognizing images of digits obtained from vehicle registration plates, which have been blurred using an unknown kernel. We model the image as an undirected graphical model over image patches in which the compatibility functions are represented as nonparametric kernel densities. The crucial feature of this work is an iterative loop that alternates between super-resolution and restoration stages. A machine-learning-based framework has been used for restoration which also models spatial zooming. Image segmentation is done by a column-variance estimation-based “dissection” algorithm. Initially, the compatibility functions are learned by nonparametric kernel density estimation, using random samples from the training data. Next, we solve the inference problem by using an extended version of the nonparametric belief propagation algorithm, in which we introduce the notion of partial messages. Finally, we recognize the super-resolved and restored images. The resulting confidence scores are used to sample from the training set to better learn the compatibility functions.