Handwritten Digit Recognition under Constrained Training Conditions
Mossad Helali, A. Alneghaimish, Irfan Ahmad · 2017
A fundamental step in handwritten digit recognition is to train a system using a large number of handwritten samples. These samples need to be accurately collected and labelled, which is a cumbersome task. In this work, we present several approaches to handwritten digit recognition in situations where little or no handwritten training data is available. We firstly study the effect of the number of training samples per digit on the recognition accuracy. We then study the effect of using machine printed digits in various font typefaces as training data on the performance of the digit recognizer. We then use some image distortion techniques to artificially generate more training data from the machine printed digits. Our final approach is to use the test set for system retraining with the classifier's transcriptions as labels. The results of our system using no labelled handwritten training data are comparable to systems using large handwritten training sets on a benchmark database of handwritten Arabic digits.