Handwritten digit recognition using hoeffding tree, decision tree and random forests — A comparative approach

The field of image computing incorporates the crucial subject of character recognition which is gaining significant importance and extensive research in the current wake of digital revolution. The goal is to find an effective software tool capable of accurately identifying the characters. The variations in a single alphabet itself and the plethora of handwriting varieties make this a challenging task. The proposed solution comprises of a series of steps for the purpose of classification. System training and testing incorporates hundred instances of handwritten digit images from MNIST Dataset. Preprocessing of the image enhances data images prior to computational processing. Input image are converted into gray scale and finally into binary. This is followed by morphological operations and the images are then converted into Comma Separated Files to be able to be used as Training and Testing Dataset in WEKA. Pattern recognition is done by Hoeffding Tree, Decision tree and Random forest methodologies to ultimately compare them on a set of benchmarks to find the most effective tool marked on a set of measures such as efficiency, effectiveness, time to perform the complete process of classification, etc. On the basis of the key parameters which include classified instances of the digits, error rate and time taken for the classification, Hoeffding tree is found to be most effective in terms of time taken to build model, precision, recall and confusion matrix. The future work requires inclusion of an extensive data set to declare the best among these approaches.