Offline Malayalam character recognition: A comparative study using multiple classifier combination techniques

Abstract

Malayalam character recognition has gained immense popularity in the past few years. The intrinsic challenges present in this domain along with the large character set of Malayalam further complicate the recognition process. Here we present a comparative evaluation of different multiple classifier combination techniques for the offline recognition of Malayalam characters. We have extracted three different features from the preprocessed character images—Density features, Run-length count and Projection profiles. These features are fed as input to three different neural networks and finally the results of these three networks were combined and evaluated using six different classifier combination methods: Max Rule, Sum Rule, Product Rule, Borda Count Rule, Majority Voting and Weighted Majority voting schemes. The best recognition accuracy of 97.67 % was attained using the Weighted Majority scheme considering top 3 results.