Classification of Hand-Written Digits Using Chordiograms

The chordiogram has recently been proposed for detection and segmentation of shapes in images. This paper evaluates the effectiveness of using chordiograms for recognizing hand written characters using the MNIST dataset. The method calculates a feature for each digit based on the geometric relationships of boundary pixels. The resultant features are used to train a support vector machine which is then used to classify a test set. A comparative study carried out for this paper shows that using boundary pixels is not as effective as calculating similar features based on the a character skeleton extracted using thinning. Character recognition error rates with skeletons as low as 2.4% are achieved. A slightly better error rate, 2.2%, can be achieved with boundary pixel chordiograms, but at the expense of making the feature vector very large. This performance is compared to the results of classifying digits based on their pixel intensities.