Handwritten Digit Recognition - Masters Thesis Summary Report

Pattern recognitionclassification of handwritten digits were performed on a random sample of 3000 digits. Each class was trained with 200 digits and tested with 100 digits. Each digit was normalized to a 32x32 matrix representation of the digit. Four methods were used to classify each digit Directional Vectors, Profiles, Curvatures, and Profile Curvatures. In addition to each method used, a wavelet transform was also performed on the digits to see if any better results could be obtained. The goal of this project was to investigate less common methods that might be useful in pattern recognition of digits while keeping the generality of these algorithms. It was not the intent of this project to base methods off certain digits and combine them together to create an algorithm used for classification each method was used on its own as a classifier. In each method the classification is done using the Mahalanobis distance function. Either the covariances are used or a number of eigenvectors based from the largest eigenvalues are used. Each method may use a different number of eigenvectors.