Channel Attention Integrated Lightweight Deep Neural Network for Bengali Handwritten Numerals Recognition

Decades of research have been devoted to the problem of handwritten character recognition regarding various languages, but the resultant existential systems have not been very effective for Bengali yet. Recognizing Bengali handwritten numerals has several distinctive obstacles, such as the diversity in writing styles, the changing forms and sizes of numerals, the fluctuating degrees of noise, and the distortion present in the images. This paper introduces a simplified, lightweight deep neural network architecture that incorporates channel attention and ensemble approach. The architecture has featured additional dropout layers and ReLU activation to enhance the network’s speed. The PUST dataset, which is the most recently available Bengali numerals dataset, was incorporated to evaluate the performance of the proposed architecture. The experimental findings demonstrated that the proposed approach recognized Bengali numerals with an accuracy of 99.00%, F1-score of 99.00%, and AUC score of 0.999. Without any rigorous image preprocessing, our proposed simplified, lightweight architecture performed the best with fewer epochs and less computational time.