Automatic Recognition of Legal Amount Words of Bank Cheques in Devanagari Script: An Approach Based on Information Fusion at Feature and Decision Level

Abstract

Legal amount word recognition is an essential and challenging task in the domain of automatic Indian bank cheque processing. Further intricacies get accumulated by inherent complexities in Devanagari script besides cursiveness present in handwriting. Due to segmentation ambiguity and variability of constituent parts present in handwritten word analytical approach is inadequate, in contrast, to the holistic paradigm, where the word is taken indivisible entity. Despite the proliferation of various feature representations, it still remains a challenge to get effective representation/description for holistic Devanagari words. In this paper, we made an attempt to exploit robust, most discriminative and computationally inexpensive Histogram of Oriented Gradients (HOG) and Local Binary Pattern (LBP) for effective characterization of Devanagari legal amount words taking into account different writing styles and cursiveness. Two models are proposed based on fusion strategies for word recognition. In the first model, LBP and HOG features are fused at feature level and in second, fused at decision level. In both models, recognition is performed by the nearest neighbour (NN) and support vector machine (SVM) classifiers. For corroboration of the results, extensive experiments have been carried out on ICDAR 2011 Devanagari Legal amount word dataset. Experimental results reveal that fusion based approaches are more robust than conventional approaches.