Text localization with hierarchical multiple feature learning

Abstract

In this paper, we focus on English text localization in natural scene images. We propose a hierarchical localization framework which goes from characters to strings to words. Different from existing methods which either bet on sophisticated hand-crafted features or rely on heavy learning models, our approach tends to design simple but effective features and learning models. In this study, we introduce a kind of two level character structure features in collaboration with the Histogram of Gradient (HOG) and the Convolutional Neural Network (CNN) features for character localization. In string localization, a nine-dimension string feature is proposed for discriminative verification after grouping characters. For the final word localization, we learn an optimal splitting strategy based on the interval cues to split strings into words. Experiments on the challenging ICDAR benchmark datasets demonstrate the effectiveness and superiority of our approach.