Deep convolutional neural networks for degraded printed kannada character recognition

Abstract

Recognition of degraded printed Kannada characters is a challenging research problem. Proposed in this paper is a deep convolutional neural network for recognition of degraded printed Kannada characters. Characters in some old Kannada texts are affected by various degradations that result in breakages and dilations of characters introducing challenges in the process of recognition. The architecture consists of three levels, the first two levels with ReLu, Max pooling layers, and the third level with just ReLu. The output of these is input to fully connected layer which performs classification of characters. Experimental analysis is carried out using 156 classes of characters each class with 100 instances. Performance is evaluated for 4 epochs with 60 iterations per epoch. Highest classification accuracy of 99.51% has been reported for 75% training.