Optimization of Deep CNN-based Bangla Sign Language Recognition using XGBoost classifier.

Deafness is one of the major health problems in Bangladesh, where almost 10% of the population falls under this category. The plight of deaf individuals to fit in with the mainstream population is significant. Unable to express themselves through sound, they have to rely on sign language to communicate.To create a more inclusive environment for this minority, we developed a model that recognizes Bengali characters and digits expressed using sign language in accordance with Bangla Sign Language (BdSL) guidelines. The model comprises two parts: a convolutional neural network that uses deep learning techniques, and an Extreme Gradient Boosting (XGBoost) classifier. Together, these components form the full architecture. The model’s execution was validated using the ‘Ishara-Lipi’ dataset, which is the first open-access digit and character dataset for BdSL. With the help of pre-processing techniques such as contrast-limited adaptive histogram equalization, using a more complex pre-trained CNN model like Inception-ResNet-v2, and optimizing the XGBoost model by using GridSearchCV, we achieved an accuracy of 86.67%, precision of 89%, recall and f1 score of 87%. Lastly, for digit recognition, we obtained an accuracy of 97.33%, precision of 98%, and recall and f1-score of 97%.