1D-LDA verses 2D-LDA in online handwriting recognition

The paper compares the performance of both one-dimensional (ID) and two-dimensional (2D) linear discriminant analysis (LDA) in recognizing online handwritten Kannada characters. The main difference between 1D-LDA and 2D-LDA is the way the data is presented to these tools for dimensionality reduction. While, the extracted features of a data sample are vertically cascaded to form a column vector for 1D-LDA, two-dimensional data is directly processed using 2D-LDA Online handwritten Kannada basic character data set is subject to experimentation to judge the performance of these tools. Writer independent experiments are conducted on training data of 3750 samples and test data of 1550 samples. The combined estimate and derivative features are fed to both 1D-LDA and 2D-LDA subspace algorithms for dimensionality reduction. With nearest neighbor as classifier, maximum average recognition accuracy of 87.4% with 1D-LDA and 87% with 2D-LDA is achieved. Experiments are also conducted to understand the dependency of both 1D-LDA and 2D-LDA for varying Eigen vectors.