A method for multi-oriented Thai text localization in natural scene images using Convolutional Neural Network

Text information in natural scene images plays an important role in many computer vision applications, such as license plate recognition, scene text recognition, and assistive text reading for visually impaired people. Unlike the existing Thai text localization methods, which use connected component analysis and rule-based techniques to locate text position, in this paper, we present a new way to locate multi-oriented Thai text based on Convolutional Neural Network (CNN). Like sliding window based object detection methods, text confidence maps are constructed by using our trained CNN text detector on multi-scaled images. The multi-scaled text confidence maps are merged to produce original input size text confidence map. By using Thai text characteristics, text line hypothesis is generated from a merged text confidence map. Finally, the post-processing techniques and Thai text characteristic analysis are performed to acquire text locations. The experimental result on our mixed Thai-English, BEST2015 Text Location Detection Contest dataset, ICDAR2003 and ICDAR2011 dataset show that our method can locate English and Thai text with promising accuracy compared to the state-of-the-art. Our proposed method won the 1st prize in BEST2015 Text Location Detection Contest organized by National Electronic and Computer Technology Center (NECTEC).