Comparing persistent homology-based classifiers for Filipino sign language recognition

Deaf or hard of hearing individuals have long been faced with problems in communication. To cope with this communication gap, numerous sign languages have been developed, one of which is the Filipino Sign Language (FSL). Despite FSL being declared as the national sign language of the Philippines, there is lack of formal implementation of policies and the gap problem continues to prevail. Sign interpreters play a crucial role in this limitation but are still insufficient in number. Hence, machine learning techniques are leveraged to automate the interpretation process of signed gestures and the field of Sign Language Recognition (SLR) is developed. This paper extends this by utilizing computational topology-based methods in performing SLR on a FSL dataset. Specifically, it aims to utilize the Persistent Homology Classification Algorithm (PHCA) in classifying or interpreting dynamic FSL gestures. Due to many distinct classes considered in this problem, this paper also aims to develop the multi-level PHCA in which this approach divides the classification task. It does so by defining categories consisting of classes and performing two-part classification, first on the categories, and second on the classes within the selected category. The performance of PHCA and multi-level PHCA using different classification metrics is evaluated. The predicting and training time of the models are also compared. Results show that both PHCA and multi-level PHCA produced satisfactory performance for SLR. It is shown further that multi-level PHCA outperformed PHCA in all setups considered, producing an accuracy of 85% for 10 classes and 60.99% on 105 classes, indicating a potential for further research.