A Comparative VGGNET and DENSENET Approaches to Recognize Malayalam Characters using Transfer Learning Techniques

Transfer Learning has become the pillar of the future when it comes to character recognition. It brought improvements in one of the south Indian languages as Malayalam. The Malayalam complex character pattern recognition from Malayalam document images is one of the recent trends in the area of recognition. Malayalam Handwritten Character Recognition, a domain under pattern recognition, has captured huge interest due to its relevance and complex type of characters. But when it comes to recognition of characters from ancient documents like manuscripts or palm leaves, the challenges get even higher. This research study focuses on the recognition of Malayalam, which is an Indian script commonly used in Southern India by using a transfer learning approach. The proposed research work aims to digitize ancient Malayalam manuscripts and other documents so that the information written in these documents can be preserved that will be used for training the system. The proposed work is a comparison recognition performance of various Deep Learning (DL) architectures in our own dataset created from various sources using the trained ancient script. In the transfer learning, the knowledge transferred from image net dataset weights to our dataset and train the system using ancient characters to recognize the new Malayalam character. The proposed models fine-tuned DENSENET and VGGNET to provide a recognition accuracy of 98.39%, 99.35% respectively for dataset containing126 classes.