Intelligent Recognition of Ancient Brahmi Characters using Transfer Learning

Character recognition is an unresolved challenge when it comes to regional languages. Researchers across the world are working on automatic document reading, Translation, indexing and ancient text collection and recognition. Similar challenges are addressed in this paper to decipher ancient brahmi characters found on various inputs. Inputs considered are handwritten brahmi characters, Printed brahmi characters and Epigraphical script images. Manual collection of dataset includes, 287 characters having 210 samples per class making the dataset big enough to train and test deep convolutional neural networks for recognition. Alexnet, Googlenet and Inception-V3networks were transfer learned for the brahmi input and Inception V3 performed well. The traditional machine learning models are trained and tested on Hu-Zernike moments and found, Transfer learning yields better results by recording huge features from the image samples.