Recognizing Filipino Sign Language with InceptionV3, LSTM, and GRU

Abstract

Sign language Recognition is a field in artificial intelligence that is gaining traction in recent years. Although there have been significant advances in both Artificial intelligence methods and Sign language recognition methods for sign languages, Filipino Sign Language is still lacking in this field. The problem lies in the dataset. With this in mind, the authors of this created a video dataset of 20 introductory Filipino Sign language words and then utilized InceptionV3 and Recurrent Neural Network (RNN), specifically Long Short Term Memory (LSTM) and Gated Recurrent Units (GRU), to develop a pipeline model that can classify the FSL signs. The author was able to create a 20-sign FSL dataset of 414 videos (at least 20 samples each). The models were trained using the dataset. Results show that GRU at 16 units performed best at this dataset with an accuracy of 86.74% Further studies on the generalizability of the architecture as well as other models trained on this dataset is recommended