MULTILINGUAL TEXT READING IN NATURAL SCENE IMAGES

Abstract

Reading the texts will greatly help to better understand the image contents. In this thesis, we address the text reading problem by solving each of the sequential components, namely, text detection, text segmentation and text recognition.

The first problem for text reading is to locate the text areas in the images. To address this, we propose a unified text detection system, i.e., Text Flow, to deal with languages of different scripts.

When the text locations have been identified, the background regions of the located texts need to be removed before conducting text recognition. We propose text segmentation algorithms from different perspectives based on color and stroke.

Once the background regions are removed, we propose convolutional CoHOG feature to recognise characters in different scripts. We also propose one Chinese and one Bengali character dataset, to advance the research in multilingual scene text recognition.