Input simplifying as an approach for improving neural network efficiency

Abstract

With the increasing popularity of smartphones and services, symbol recognition becomes a challenging task in terms of computational capacity. To our best knowledge, existing methods have focused on effective and fast neural networks architectures, including the ones which deal with the graph symbol representation. In this paper, we propose to optimize the neural networks input rather than the architecture. We compare the performance of several existing graph architectures in terms of accuracy, learning and training time using the advanced skeleton symbol representation. It comprises the inner symbol structure and strokes width patterns. We show the usefulness of this representation demonstrating significant reduction of training time without noticeable accuracy degradation. This makes our approach the worthy replacement of conventional graph representations in symbol recognition tasks.