An intelligent fusion-based transfer learning model with artificial protozoa optimiser for enhancing gesture recognition to aid visually impaired people

Generally, the interaction of gestures presents a set of benefits to persons with disabilities, from improving motor, social, and cognitive skills to delivering a secure and controlled atmosphere for engaging in real-world scenarios. Automatic detection methods, such as hand gesture recognition (GR), are among the most active research fields. Hand gestures, particularly in the form of sign language (SL), are one of the effective modes of non-verbal interaction and a stimulating area of research because they can simplify communication and serve as a standard method of communication, employed in a variety of applications. Particularly, GR has received significant attention in the area of Human-computer interaction. Recently, artificial intelligence (AI) technology has been extensively used by scholars to achieve GR and accurate computer vision. Deep learning (DL) is a rapidly evolving technology that aims to simplify a method by which deaf individuals can connect with others. This study presents an Artificial Protozoa Optimiser-Based Fusion of Transfer Learning Models for Enhancing Gesture Recognition (APOFTLM-EGR) model. The APOFTLM-EGR model aims to enhance GR for visually impaired individuals. The image pre-processing stage applies Wiener filtering (WF) to eliminate the redundant or unwanted noise in the input image data. Furthermore, the VGG16, InceptionV3, and ResNet-50 fusion models perform the feature extraction process. The stacked sparse autoencoder (SSAE) technique is employed for the GR process. Finally, the artificial protozoa optimizer (APO) technique optimally adjusts the hyperparameter values of the SSAE method, resulting in improved detection performance. The efficiency of the APOFTLM-EGR method is validated by comprehensive studies using the Indian SL dataset. The experimental validation of the APOFTLM-EGR method delivered a superior accuracy value of 99.46%, outperforming existing models.