GridMask Based Data Augmentation for Bengali Handwritten Grapheme Classification

Abstract

In this paper, we describe the deep learning-based Bengali handwritten grapheme classification. Specifically, our recognition approach is based on the convolutional neural networks (CNNs) as deep CNNs have achieved splendid performance on many different visual recognition tasks. Moreover, we employ GridMask-based data augmentation to improve the recognition performance further. We compare the GridMask-based data augmentation with conventional data augmentations (such as flip, rotation, mixup) on three widely-used CNN architectures: ResNet101, DenseNet169 and EfficientNet B0. Extensive experiments demonstrate GridMask can utilize the information removal to improve the robustness of the neural networks, and the boost of hierarchical macro-averaged recall on the validation set suggest that GridMask data augmentation can be efficiently used for the Bengali handwritten grapheme analysis without any prior grapheme segmentation.