Recursive Feature Elimination for Improving Learning Points on Hand-Sign Recognition

Hand gestures and poses allow us to perform non-verbal communication. Sign language is becoming more important with the increase in the number of deaf and hard-of-hearing communities. However, learning to understand sign language is very difficult and also time consuming. Researchers are still trying to find a better way to understand sign language using the help of technology. The accuracy of most hand-sign detection methods still needs to be improved for real-life usage. In this study, Mediapipe is used for hand feature extraction. Mediapipe can extract 21 hand landmarks from a hand image. Hand-pose detection using hand landmarks is chosen since it reduces the interference from the image background and uses fewer parameters compared to traditional hand-sign classification using pixel-based features and CNN. The Recursive Feature Elimination (RFE) method, using a novel distance from the hand landmark to the palm centroid, is proposed for feature selection to improve the accuracy of digit hand-sign detection. We used three different datasets in this research to train models with a different number of features, including the original 21 features, 15 features, and 10 features. A fourth dataset was used to evaluate the performance of these trained models. The fourth dataset is not used to train any model. The result of this study shows that removing the non-essential hand landmarks can improve the accuracy of the models in detecting digit hand signs. Models trained using fewer features have higher accuracy than models trained using the original 21 features. The model trained with 10 features also shows better accuracy than other models trained using 21 features and 15 features.