Malayalam Handwritten Character Recognition using Transfer Learning and Fine Tuning of Deep Convolutional Neural Networks

In the digitization of Malayalam handwritten documents, recognition of handwritten characters is a difficult task. This is due to the non availability of a labeled benchmark Malayalam handwritten character dataset. The state of the art technique using deep convolutional neural networks demands large amount of labeled dataset. Therefore, this paper aims to develop a pre-trained convolutional neural network (CNN) model for recognizing Malayalam handwritten characters using small sized dataset. Two approaches namely transfer learning and fine tuning of pre-trained Deep Convolutional Neural Network (DCNN) architecture ResNet50 are used to develop models for recognizing Malayalam handwritten characters. Model design is optimized by varying parameters like learning rate, batch size and optimization algorithm. From the experiments, it is found that highest testing accuracy of 78.05% is obtained for the model using fine tuning approach when it is trained with a batch size of 16 using RMSProp optimization algorithm and a learning rate of 0.000001. A testing accuracy of 78.05% is obtained with ResNet50 for binary images even though ResNet50 is pre-trained using colour images.