Robust scene text recognition: Using manifold regularized Twin-Support Vector Machine

Abstract

Text recognition in natural scene images endures as a challenging problem, attributable to a highly variable appearance in an unconstrained environment. The proposed work is novel of its kind, which improved the generalization of the Twin Support Vector Machine (T-SVM) by manifold regularization, that was further extended to validate and recognize the text in the natural scene images. The multi-class T-SVM is augmented with ambient regularizer term and intrinsic regularizer term, assisting to form a smooth function for the model. Text comprehension from natural scene includes the localization, recognition, and reconstruction of the text. The work includes an additional module, revalidation that discards the false positives from the text objects detected during the localization phase. Then in recognition phase, each letter from the pool of text objects is identified for the appropriate class and provided as output to the text construction phase, which builds the text using the pool of coordinates associated with the object. The model is evaluated against traditional methods like Support Vector Machine (SVM), T-SVM, LST-SVM(Least Square Twin Support Vector Machine) and also with other research of the same kind and shows the accuracy of 84.91% with ICDAR 2015, 84.21% with MSRA 500 and 86.21% with SVT. The analyses of the results show that the model was capable of recognizing most of the characters from the image along with consummate accuracy level.

Keywords

• Machine learning
• Manifold regularization
• Multi-class
• Scene text detection
• Scene text recognition
• Twin Support Vector Machine