IncepForest: A Novel Approach for Sign Language Recognition with Transfer and Ensemble Learning

Recent advances in the field of the human-computer interface have sparked a surge of interest in hand gesture detection due to its inherent flexibility and user-friendliness. Gesture recognition techniques have proven invaluable in creating systems that facilitate seamless information exchange among impaired individuals or device operations. As ubiquitous computing continues to progress, traditional user interaction methods like keyboards, mice, and pens are proving inadequate, thus necessitating more natural input methods such as direct hand input. In this study, the application of transfer learning methods for hand gesture detection in Indian Sign Language (ISL) and Bengali Sign Language (BSL) is investigated. To this end, several transfer learning techniques including InceptionResNetV2, VGG16, InceptionV3, and EfficientNet were employed on both datasets. Results demonstrate that InceptionResNetV2 outperformed the other transfer learning approaches, achieving an accuracy of 99.72 % for ISL and 91.00% for BSL. To further enhance the prediction accuracy of hand gesture recognition, several hybrid approaches were implemented by integrating InceptionResNetV2 with ensemble learning models like Bagging, Boosting, and Random Forest Algorithms. This study proposes a novel hybrid model, named IncepForest, that combines InceptionResNetV2 with Random Forest and yields the highest accuracy of 96.00% for BSL.