Scene Text Recognition using part-based TSM and Soft Output Viterbi Algorithm

摘要

In recent years, detecting text in natural is an emerging problem and gained increasing attention from the computer vision community. Detecting text in natural scene text image is an important step for number of applications such as computerized aid for visually impaired, automatic sign reading, language translation and navigation. Natural scene text image may contain complex background and increased the challenges for the research community in detection and recognition. The novel scene text recognition method uses Part-based tree-structure model to improve the performance of the system by modeling each category of character to detect and recognize simultaneously. For word recognition, Soft Output Viterbi Algorithm was used to improve the measure of reliability in hard bit decision of the Viterbi algorithm. It maximizes the character sequence posterior probability using Bayesian decision view and n-gram model. To evaluate the performance of the system Char74k dataset was used for character detection and two more datasets ICDAR2003 and SVT was used for word recognition.