Character Recognition in Natural Scene Images

Applications like content-based image indexing, automatic geocoding of businesses and real-time robotic navigation have generated research interest in the problem of text reading from natural images. This paper tackles the problem of character recognition which constitutes the last phase of the text detection and extraction pipeline. We employ the stroke width transform[12] used for the text detection phase, and address the problem of character recognition by training KNNs, Random Forest and Neural Networks on the Char74 and ICDAR2003 datasets of images in-the-wild, containing English characters from natural images. We also compare the performance of our character classifier from the currently best-performing state-of-the-art OCR [11] and demonstrate that both perform competitively.