Improving Isolated Digit Recognition Using a Combination of Multiple Features

Abstract

Abstract — This paper investigates the combination of different statistical and structural features for recognition of isolated handwritten digits, a classical pattern recognition problem. The objective of this study is to improve the recognition rates by combining different representations of non-normalized handwritten digits. These features include some global statistics, moments, profile and projection based features and features computed from the contour and skeleton of the digits. Some of these features are extracted from the complete image of digit while others are extracted from different regions of the image by first applying a uniform grid sampling to the image. Classification is carried out using one-against-all SVM. The experiments conducted on the CVL Single Digit Database realized high recognition rates which are comparable to state-of-the-art methods on this subject. Keywords—Isolated handwritten digits; feature combination, Support Vector Machine. I. I NTRODUCTION Handwriting recognition has been the premier research problem of the document analysis and recognition community for over three decades now. The sub problems in handwriting recognition mainly include line, word or character level segmentation, recognition of isolated characters, words, or complete lines/paragraphs and recognition of numerical strings and isolated digits. Among these different modalities of handwriting recognition, this research focuses on recognition of isolated digits, a classical pattern recognition problem that offers a wide range of applications. Unlike alphabet, the ten glyphs of the most commonly used Western Arabic numerals are shared by many scripts and languages around the world making them globally acceptable. The main challenges in handwritten digit recognition arise from variations in size, shape, slant, and most importantly, the differences in the writing styles of individuals. With the recent advancements in image analysis and pattern classification, sophisticated digit recognition systems have been proposed which aim to enhance the overall recognition performance by improving the feature extraction or/and classification techniques used. Some of the studies aim to improve the classification performance by using a combination of multiple classifiers while others aim to combine multiple features and select the most pertinent and optimum set of features for this problem. In this paper, we are interested in enhancing the feature extraction step for isolated digit recognition used for avoiding digit normalization. The idea is to find a combination of multiple features which improves the overall recognition rates by minimizing the intra-class variability and maximizing inter-class variability [4, 13, 23], the most desirable requirement of any pattern recognition system. Over the years, various handwritten isolated digit recognition systems reporting high recognition rates have been proposed. Most of these systems have been evaluated on the widely used MNIST database. Recently, the handwritten digit recognition competition [19] held in conjunction with ICDAR 2013 also provided a platform for comparison of state-of-the-art digit recognition techniques under the same experimental conditions. Among significant contributions to digit recognition, authors in [31] present a comprehensive comparison of different classification algorithms on the recognition task. Heutte et al. [15] proposed a combination of seven different features to feed a linear discrimination based classifier. Dong et al. [12] extracted a set of gradient features while Teow and Loe [16] computed linearly separable features from the MNIST database and applied triowise linear support vector machine with soft voting for classification. Belongie et al. [27] developed a novel similarity measure by finding the correspondences between points in two shapes and estimating an aligning transform. The proposed matching technique achieved high recognition rates when applied to digit recognition. In another notable contribution, Lauer et al. [7] proposed a trainable feature extractor based on LeNet5 neural network architecture. Classification carried out using Support Vector