A new approach for upscaling document images for improving their quality

One of the issues faced by optical character recognition (OCR) softwares is the input document images being not of good quality. So research into the methods of enhancing the document images, before presenting them to OCR softwares, is of utmost importance. The objective is to demonstrate a method of generating a high resolution document image, given a low resolution image. We propose a new method for improving the spatial resolution of document images. Here, we have built a deep neural network based model that utilizes the traditional interpolation methods, takes the best features from them and reconstructs a high resolution image from these features. This is achieved using a convolutional neural network (CNN). The CNN learns a high resolution patch from a corresponding low resolution patch, as a weighted non-linear combination of the outputs of different interpolation techniques. We call our technique as nonlinear fusion of multiple interpolations (NFMI). The NFMI method ensures that the model learns only the best features that can be extracted from all the interpolation techniques combined together. The use of traditional interpolation methods makes sure that the NFMI technique is not computationally expensive. Results on test images show a relative improvement of 54% in word recognition accuracy by OCR over the best interpolation technique for doubling the spatial resolution and 33% for quadrupling the resolution.