In this paper we present a resizing neural network for edge and detail preserving image interpolation. The multilayer neural network is trained by using pairs of high resolution and low resolution imagery. The high resolution is an 8-bit image scanned at 600 dpi. The low resolution image (300 dpi) is either a processed version of the high resolution image, or it is scanned independently. pixels are extracted from the low (high) resolution image and are used as inputs to the neural networks. The interpolated pixels obtained as output are compared with the high (low) resolution pixels after enhancement and the error is used to train the neural network.
Mohamed N. Ahmed, Brian E. Cooper, Shaun T. Love, "Document Resizing Using a Multi-Layer Neural Network" in Proc. IS&T Int'l Conf. on Digital Printing Technologies (NIP17), 2001, pp 792 - 796, https://doi.org/10.2352/ISSN.2169-4451.2001.17.1.art00080_2