A B S T R A C T
Sign Language Recognition (SLR) is used to communicate between deaf and normal people or among hearing-impaired communities. The rapid development of Artificial Intelligence technologies can be exploited as a medium to strengthen these connections. Many studies have
been conducted to indicate their interest and usefulness. The differences between this study and previous research are as follows: first, increasing the noise robustness and noise
robustness analysis because noise can always appear in real applications; Secondly, an architecture is proposed to extend the dataset without merging and retraining all datasets. Our
study proposes using a Convolutional Neural Network (CNN) as a feature extractor; different potential models are compared for the effectiveness and complement of feature extractors. Feature normalization was analyzed before being fed to the ensemble classifier. Moreover, our study compared Support Vector Machine (SVM), random forest, and bagging SVM as optimized candidates for the ensemble classifier. Our study shows that the proposed micro-block provides better accuracy than the reference SLRNet-8 model in dataset D1
with some noisy environments, and the proposed transdataset ensemble architecture is also better than applying the traditional transfer learning technique with reference SLRNet-8 model when validating with either dataset D1 or D2 in our setting environment.
Keywords—sign language, convolutional neural network, ensemble classifier, model fusion, random forest, feature extractor