SVM and MLP Based Segmentation and Recognition of Text from Scene Images Through an Effective Binarization Scheme

Abstract

Text Binarization from scene images plays a crucial task for any text segmentation scheme and therefore in the OCR performance. So, an effective image binarization method is required for text segmentation and recognition tasks. This work describes an effective image binarization scheme for segmentation and recognition task of text from images. To binarize the image, Canny’s edge information is incorporated into Otsu’s method. It generates numerous components which are analyzed for segmentation of probable text components. Further, a few features are considered for classification of text and non-text. For this problem, SVM is considered. For training SVM classifier, information from ground-truth images of text and our own made non-text components are used. Finally, Multilayer Perceptron (MLP) is used for recognition of the text symbols. The MLP classifier is trained using 13496 samples of 39 classes. We tested our schemes on the publicly available ICDAR Born Digital data set. The outcomes are quite acceptable.