RAS-bengaliNet: An Explainable Lightweight Convolutional Neural Network for Handwritten Bangla Compound Characters Recognition

Bengali is the fifth most widely spoken native language and the seventh most widely spoken language worldwide, deserving precise attention in the domain of character recognition.However, the complexities of Bengali characters have frequently presented difficulties, limiting significant progress in comparison to other languages.This study introduces a novel lightweight convolutional neural network (CNN) architecture for recognizing Bengali handwritten compound characters.A dataset comprising over 8000 compound characters from 119 categories was meticulously curated, sourced from diverse volunteers across Bangladesh, establishing a robust foundation for our research.Our proposed architecture exhibited superior performance in terms of classification accuracy and computation time when compared to state-of-the-art CNN models such as VGG16, ResNet50, and MobileNetV2.Through thorough experimentation, our model demonstrated exceptional performance, surpassing previous benchmarks on our dataset.Furthermore, we conducted a comprehensive evaluation using well-known datasets including BanglaLekha-Isolated, CMATERdb 3.1.3.1 and 3.1.3.3, and Ekush, which confirmed the effectiveness of our approach.Specifically, our methodology achieved remarkable accuracies of 99.59%, 99.93%, 99.87%, and 99.94% on these datasets, respectively.Notably, on our proprietary compound character dataset, our proposed model achieved a perfect accuracy of 100.00%, further emphasizing the efficacy of our approach.Additionally, we utilized Gradient Class Activation Mapping (Grad-CAM) to identify the crucial regions responsible for decision-making.Therefore, this study advances