The Recognition of Tibetan Handwritten Numbers Based on Federated Learning

Abstract

Numeral recognition is closely related to individual lives, involving postal codes and bank checks. In recent years, despite many researches focusing on handwritten recognition, there were few on the identification of Tibetan handwritten numerals; and related to privacy protection. This paper proposed a recognition system based on lightweight CNN and federated learning, aiming to reduce the total calculating resource consumption and secure sensitive information. Besides, pre-processed TibetanMNIST dataset were adopted as the training samples for case study. This cases eventually obtained nearly 96% accuracy, and the expected time required to process a single image was approximately 0.017ms. The proposed system can be used for the recognition of Tibetan handwritten numbers.