Handwritten Digit Recognition Using Bayesian ResNet

Abstract

The problem of handwritten digit recognition has seen various developments in the recent times, especially in neural network domain. The methods based on neural network work quite effectively for the seen classes of data by providing deterministic results. However, these methods tend to behave in similar fashion even for unseen class of data. For example, a neural network trained on English language digits will give a deterministic prediction even when tested on digits of other languages. Hence, it is required to predict uncertainty for such methods in this scenario. In this paper, we employ Bayesian inference into the existing ResNet18 framework to bring out uncertainty for handwritten digit recognition when there is a new class of test digit. We term the new architecture as B-ResNet. The novel B-ResNet is first of its kind to be investigated for the handwritten digit recognition. Various experiments on datasets of English, Devanagari, Gujarati, Bengali digits and their all possible combinations demonstrate the efficiency and performance of the B-ResNet for hand written digit recognition.

Self-Adaptive Hybridized Lion Optimization Algorithm With Transfer Learning for Ancient Tamil Character Recognition in Stone Inscriptions

“…The detection system's final step is to categorise discovered objects into three categories namely noise, Balinese script, and hole. [29] have tried using Modified ResNet-18 algorithm for recognizing Bangla handwritten characters. ResNet model is chosen as the underlying architecture since it produces better results than other compared architecture models in different applications and the same is proved in this implementation too.…”