Citation

Lim, Alvin Fang Chuen (2024) In-Air Hand Gesture signature recognition using Bi-Directional Recurrent Neural Network with deep convolutional features. Masters thesis, Multimedia University.

Full text not available from this repository.

Abstract

A hand signature is a type of behavioral biometric. It can be represented in the form of a handwritten name or symbol that serves as proof of identity. Hand signature recognition involves classifying signatures to determine to whom they belong. Most commercially available hand signature recognition systems rely on touch-based handwriting, as it is cost-effective and has proven reliable. However, this approach is found to be unhygienic, as the acquisition devices need to be shared. Thus, this thesis proposes a contactless hand signature recognition method utilizing deep learning, where there is no direct contact with the acquisition devices. The In-air Hand Gesture Signature (iHGS) database is used to evaluate the proposed method. The raw samples are not suitable for training the model. Pre-processing is employed to remove noise from the samples, which includes palm detection and segmentation. A custom feature generation algorithm is proposed for 2-dimensional and 3-dimensional features, termed iHGS-MHI and iHGS-MHI-BLOCKS, respectively. Preliminary experiments were conducted to compare the performance of pre-trained model architectures and the proposed CNN architecture. The objective is to identify and select the architecture that yields the highest accuracy with minimal computation time for training as the convolution layer in the integrated architecture of Convolutional Recurrent Neural Networks (C-RNN). The proposed CNN-A architecture, which achieved the highest accuracy of 95.10% and an average training time of 1 minute and 42 seconds, has been selected for the C-RNN. Three variants of C-RNN were proposed: ConvLSTM, ConvBiLSTM, and ConvGRU. Subsequent experimental results show that all variants of C-RNN are computationally efficient and achieve high accuracy with 128 hidden units, all obtaining accuracy within the range of 97% and an average training time of 3 minutes. Notably, the highest accuracy achieved was 98.10% on ConvBiLSTM with 512 hidden units, with a training time of 7 minutes.

Item Type:	Thesis (Masters)
Additional Information:	Call No.: QA76.87 .L56 2024
Uncontrolled Keywords:	Neural networks (Computer science)
Subjects:	Q Science > QA Mathematics > QA71-90 Instruments and machines > QA75-76.95 Calculating machines
Divisions:	Faculty of Information Science and Technology (FIST)
Depositing User:	Ms Nurul Iqtiani Ahmad
Date Deposited:	03 Feb 2025 03:47
Last Modified:	03 Feb 2025 03:47
URII:	http://shdl.mmu.edu.my/id/eprint/13343

Downloads

Edit (login required)