Abstract

In the problem of handwritten digit classification, difficulties are encountered when there is ambiguity among the digits to be classified. It is desirable to detect this confusion and either reject the classification or attempt to make a better decision using post-processing methods. In the proposed method, a cascading neural network model is used to do the latter. Each level contains an ensemble of neural networks trained on different features. This generates classifiers that complement each other and help identify samples that are difficult to classify. Experiments were done on the MNIST database. This database has 60000 training images and 10000 test images that contain segmented handwritten digits. The results from the experiment show an improvement in the classification accuracy with the addition of each level of neural networks. Out of the 10000 test images 2206 of the samples were rejected from the cascading neural network model and were sent to post-processing. Among the 7794 of the accepted samples not sent to post-processing, only 52 were falsely classified. The overall classification rate of the system, including post-processing is 96.58%