Text localization and recognition in natural scene images

Text localization and recognition (text spotting) in natural scene images is an interesting task that finds many practical applications. Algorithms for text spotting may be used in helping visually impaired subjects during navigation in unknown environments; building autonomous driving systems that automatically avoid collisions with pedestrians or automatically identify speed limits and warn the driver about possible infractions that are being committed; and to ease or solve some tedious and repetitive data entry tasks that are still manually carried out by humans. While Optical Character Recognition (OCR) from scanned documents is a solved problem, the same cannot be said for text spotting in natural images. In fact, this latest class of images contains plenty of difficult situations that algorithms for text spotting need to deal with in order to reach acceptable recognition rates. During my PhD research I focused my studies on the development of novel systems for text localization and recognition in natural scene images. The two main works that I have presented during these three years of PhD studies are presented in this thesis: (i) in my first work I propose a hybrid system which exploits the key ideas of region-based and connected components (CC)-based text localization approaches to localize uncommon fonts and writings in natural images; (ii) in my second work I describe a novel deep-based system which exploits Convolutional Neural Networks and enhanced stable CC to achieve good text spotting results on challenging data sets. During the development of both these methods, my focus has always been on maintaining an acceptable computational complexity and a high reproducibility of the achieved results.