The Validation of Gesture-Based Datasets for Arabic Sign Language
Miada Almasre, Hana Al-Nuaim · 2018
The objective of this research is to validate data captured by users gesturing Arabic sign language (ArSL) letters to be interpreted correctly when signed by other users. A supervised machine learning hand-gesturing prototype was developed to capture data, using Kinect with an LMC. A Dataset of 1400 gestured ArSL letters was generated and a classifier algorithm was used to recognize each letter. Tests proved validity and usefulness of the dataset as a training set. The dispersion of the data was due to the values wide range. The data correlation tests show a strong association with the data of each letter captured from different users. Using the Support Vector Machine (SVM) classifier achieved 90% overall accuracy. The SVM results were tested through the Receiver Operating Curve (ROC) drawn for each letter. The area under the curve was calculated for individual letters and overall for all letters with 0.9075 accuracy recognition rate.