Bangladeshi Indigenous Group Identification using Deep Learning Approach

Abstract

This study presents a general investigation into the identification of two prominent indigenous groups in Bangladesh, namely the “Chakma (চাকমা)” and “Marma (মারমা)”, by the application of convolutional neural networks (CNNs). This work used some Python-based CNN algorithm and an accompanying web application capable of nearly accurately recognizing indigenous individuals. The research dataset comprised 664 facial images, with 328 belonging to the Chakma group and 336 to the Marma group. The primary objective was to evaluate and identify through comparing the performance of various CNN models, including ResNet50, VGG16, InceptionV3, and DenseNet121, in processing these image datasets. Among the tested models, ResNet50 emerged as the most proficient, achieving a Training Accuracy of 95.85% and a Test Accuracy of 95.52%. These results underscore the exceptional efficiency and adaptability of ResNet50 for the task of indigenous group identification. On the other hand, DenseNet121 got the highest accuracy in training is 97.92%. But test accuracy is lower than ResNet50, that is 86.57%. Beyond the technical aspects, the project also explores the potential applications of indigenous group identification, including preservation of cultural heritage, anthropological research, social justice and human rights advocacy, healthcare and public services customization, educational representation, forensic identification and customized services and products development. By shedding light on the utilization of deep learning techniques for indigenous group identification, this research contributes to both the academic understanding of computer vision applications and the practical implications for various societal sectors. Moreover, the developed web application provides a tangible tool for the recognition and acknowledgment of Bangladesh's diverse indigenous communities.