Towards scene text recognition with genetic programming

Recognizing text captured in a photograph, or scene text, remains an unsolved problem in computer vision. Conventional methods require a complex multi-step process to incorporate a pipeline of manually constructed algorithms. In contrast this research presents a single step framework which is based on Genetic Programming (GP). With a suitable methodology, character recognition programs can be automatically generated which are capable of handling common challenges in scene text including gradated foreground and background, low contrast, variations in size and font, without specific components designed for these challenges. Furthermore, the solutions evolved by GP are capable of handling a degree of blur and rotation without adding any extra mechanisms into the proposed GP framework. We also show that GP programs trained on synthetic images can recognize characters in real scene text images, which indicates that some genuine characteristics of text have been captured by these programs. This research lays a foundation toward a scene text recognition method which does not rely on complex preprocessing, localisation and feature extraction processes. This work shows that it is possible to evolve character recognition programs with minimal human effort.