A Deep Learning Framework for Recognizing Vertical Texts in Natural Scene

Abstract

Text detection and recognition of various text orientation in natural scene images is important to understand the surrounding. Various research has been carried out on detecting horizontal text, oriented text, curved text and vertical text. However, there are limited research carried out on detecting and recognizing vertical scene text. Vertical scene text can be differentiated into top-to-bottom vertical text, bottom-to-top vertical text and horizontal-stacked vertical text. Apart from detecting horizontal text and oriented text, the model proposed in this research detects all types of vertical scene text. In this paper, the modelling of a hybrid vertical scene texts recognition is proposed. The model contains Multidirectional Text Detector (MTD) for detecting text region in the scene image, Two-dimensional Decoder (2DD) for recognizing texts in the text region, while Vertical Oriented Network (VON) is applied to determine the types of vertical texts.