Gesture similarity index reduction using GML classifier in hybrid pretrained 3D-CNN framework

Abstract

Self-e-learning platform of gesture recognition system (GRS) is failed to retrain maximum accuracy of gesture recognition via targeted special people observation. This occurs when there is a constant flow of action come gestures and similarity index gestures in the words. Some of the letters in Indian Sign Language (ISL) have similarity index gestures (like e, f, m, and n), while others use action come gestures (say h, j, y). Many new and updated gestures for the English alphabet have been introduced ISL recently, including the similarity index gesture and the action come gesture. In spite of this, GRS is unable to adjust to the continuous changes that occurred over the poses of gestures. Global Gestures Synchronous is required for better classification of similarity index gesture and action come gesture in special people e-learning education platform. In this work, we propose a hybrid pretrained 3D convolutional neural network (HP 3D-CNN) framework to distinguish between similarity index gestures and action come gestures with the help of a gradient-based machine learning (GML) classifier. This research investigates a Probability of Concurrent Gestures (PPCG) model for estimating the likelihood that two gestures will occur simultaneously based on their similarity index and the direction of action. We break down the difficult joint prediction problem into simpler marginal sub-prediction problems. Our GML model efficiently classifies both similarity index gesture and action come gesture in the joint prediction space in which the marginal prediction model and the conditional prediction model are processed. The results show that the proposed model successfully combines the characteristics of the interacting agents to increase the likelihood of making accurate predictions of their joint action. The acquired positive feature characteristic is then used to categorise gestures in an HP 3D-CNN framework. Finally, we use the industry standard ISL benchmark database to assess the performance of our proposed GML classifier. We have shown that our HP 3D-CNN framework outperforms existing state-of-the-art methods on a variety of metrics, including Accuracy, F-measure, Recall, Precision, Dice, and Jaccard indexes.