Neural Approach for the Magnification of Low-Resolution Document Images
Zakia Kezzoula, Soumia Faouci, Djamel Gaceb · 2018
This paper proposes a new super-resolution method for degraded images of documents, captured in low resolution by mobile device. This is an improvement of a non-linear existed method but limited by its high complexity and low quality on degraded images, caused in general by the JPEG compression. On this category of images, it is necessary to increase the perimeter of the local analysis in order to obtain a better visual rendering with a reduced complexity. These constraints have been the target of our contribution, which aims at the linearization of the existed approach by the use of a bio-inspired approach based on multilayer perceptron neural networks. We have demonstrated that they are able to learn the mechanism of a super-resolution approach and make it possible to extend it, essential for its quality. This is a new alternative to the conventional use of neural approaches in image magnification.