Feature Extraction based on DCT for Handwritten Digit Recognition

Feature extraction is a crucial and challenging step in many pattern recognition problems and especially in handwritten digit recognition applications. However, the extraction of the most informative features with highly discriminatory ability to improve the classification accuracy and reduce complexity remains one of the most important problems for this task. This work investigates the effectiveness of four feature extraction approaches based on Discrete Cosine Transform (DCT) to capture discriminative features for handwritten Digits recognition. These approaches are: DCT upper left corner (ULC) coefficients, DCT zigzag coefficients, block based DCT ULC coefficients and block based DCT zigzag coefficients. The coefficients of each DCT variant are used as input data for Support Vector Machine (SVM) classifier to evaluate their performances. The objective of this work is to identify the optimal feature extraction approach that speed up the learning algorithms while maximizing the classification accuracy. We used the well known MNIST database in two variants: raw data and preprocessed data that dismisses the non-informative region of the observations in the dataset. The results have been analyzed and compared in terms of classification accuracy and reduction rate and the findings have demonstrated that the block based DCT zigzag feature extraction yields a superior performance than its counterparts.