Handwritten Character Recognition Based on Adabelief Optimized Convolutional Neural Network

Neural Network plays an important role in image classification. It uses Convolutional neural networks (CNN) which is successful on image classification. The reason for using CNN is performs better on features than any other. Alphanumeric character recognition is a very useful and powerful system to translate human written characters, extract text from documents, detect traffic signals, etc. The recognition technologies which are already in use are not very reliable and it needs to be improved further in the accuracy and efficiency. Here both alphabets as well as numbers are recognized from the EMNIST dataset (Extended Modified Institute of Standards and Technology database) which consists of alphanumeric characters. First the required dataset is imported, and it is preprocessed. Then they are trained, and models are created. Those models are run over the test set. Then the input image is fed. Binarization is done followed by feature extraction. The analyzation is done with two optimizers – Adam and Adabelief. This article discussed about few models by analyzing the validation and training accuracy and losses. Further, different EPOCH values are set for the models and the validation accuracy obtained for Adam is 97% and Adabelief is 98.5 and the losses are 8% and 4% respectively. Hence, Adabelief performs better than Adam in terms different performance metrics such as training duration, accuracy and loss.