Preprocessing and Augmentation Techniques to Improve Hand Gesture Recognition Accuracy

Abstract

HGR systems are geometrically rising in Human-Computer Interaction (HCI) applications such as sign language translation, virtual and augmented reality, robotic control, and smart environments. But they tend to perform more poorly in the presence of variation in lighting, orientations of the hands, occlusions, and complexity of the backgrounds that result in low accuracy and generalization. In the given work, the idea is to study how systematic preprocessing and data augmentation schemes can improve the resilience and accuracy of models of geometric gestures recognition. Some of the preprocessing steps include grayscale conversion, background suppression, histogram equalization, noise removal, and morphological operations to ensure that the input features are standard and to improve gesture segmentation. To increase the diversity of data in the dataset and enhance model generalization, augmentation strategies such as geometric transformations, random noise injection, color jittering, and techniques of generating synthetic data through the use of Generative Adversarial Networks (GANs) are also used. An experimental analysis of convolutional neural networks trained and tested on a benchmark dataset indicates that preprocessing on its own can improve validation accuracy by more than 10%. Combining preprocessing and augmentation can further improve cross-dataset accuracy by about 18% on average. All these findings take into account how crucial data preparation is in realising high-performance HGR systems. The paper also indicates how adaptive augmentation policies and real-time AI-guided preprocessing pipeline ought to be investigated in future studies to enable real world use of deployment in dynamic environments in a scaled fashion.