Off-line handwritten numeral recognition using the orthogonal Gaussian mixture model

The performance of the statistical approach to off-line handwritten numeral recognition is improved by using the Gaussian mixture model (GMM) to approximate an arbitrary class conditional probability density. For simplification, the GMM is commonly assumed to have diagonal covariance matrixes. Statistical correlation of the feature of handwritten numerals requires a large number of mixture components to obtain a good approximation. To solve this problem, the feature vectors are first transformed to the space spanned by the eigenvectors of the covariance matrix to reduce the correlation among the elements. The GMM is then applied to the transformed feature vectors. This GMM is defined as the orthogonal Gaussian mixture model (OGMM) which gives a better approximation than GMM with the same number of mixture components. The OGMM parameters can be estimated by EM algorithm. The algorithm effectiveness is demonstrated by applying it to the NIST (National Institute of Standards and Technology) database.