Maximizing Generalization: The Effect of Different Augmentation Techniques on Lightweight Vision Transformer for Bengali Character Classification

Deep Learning Models have proven to be highly effective in computer vision, with deep convolutional neural networks achieving impressive results on various Computer Vision tasks. However, they rely strongly on having large datasets to avoid overfitting. When a model learns features with either low or high variance, it leads underfitting or overfitting on the training data. Unfortunately, wide-ranging dataset might not exist in many fields such as resource limited languages like Bengali. In our experiment, a series of the tests were conducted in the field of Image data augmentation as an approach to addressing the limited data problem for Bengali handwritten characters. It also provides an in-depth analysis of the performance of these augmentation techniques. Data augmentation refers to a set of techniques or procedures applied to data to make them larger in size and more precise, which make them more suitable for training deep learning models. The image augmentation techniques evaluated in this study include CLAHE, Random Rotation, Random Affine, Color Jitter, and thier combinations. The study also explores the use of augmentation methods in lightweight model like EfficientViT. Among different augmentation methods, we found combination of Random Affine and Color Jitter produced the best accuracy in Ekush [1] and AIBangla [2] datasets, with accuracy of 97.48% and 97.57% respectively. This combination outperformed all other combinations, with or without augmentation techniques. This analysis presents a thorough examination of the impact of image data augmentation in resource-scarce languages, particularly in the context of Bengali handwritten characters recognition using lightweight model.