An Enhanced Hybrid Model Based on CNN and BiLSTM for Identifying Individuals via Handwriting Analysis

Abstract

Handwriting is a unique and significant human feature that distinguishes them from one another. There are many researchers have endeavored to develop writing recognition systems utilizing specific signatures or symbols for person identification through verification. However, such systems are susceptible to forgery, posing security risks. In response to these challenges, we propose an innovative hybrid technique for individual identification based on independent handwriting, eliminating the reliance on specific signatures or symbols. In response to these challenges, we propose an innovative hybrid technique for individual identification based on independent handwriting, eliminating the reliance on specific signatures or symbols. Our innovative method is intricately designed, encompassing five distinct phases: data collection, preprocessing, feature extraction, significant feature selection, and classification. One key advancement lies in the creation of a novel dataset specifically tailored for Bengali handwriting (BHW), setting the foundation for our comprehensive approach. Post-preprocessing, we embarked on an exhaustive feature extraction process, encompassing integration with kinematic, statistical, spatial, and composite features. This meticulous amalgamation resulted in a robust set of 91 features. To enhance the efficiency of our system, we employed an analysis of variance (ANOVA) F test and mutual information scores approach, meticulously selecting the most pertinent features. In the identification phase, we harnessed the power of cutting-edge deep learning models, notably the Convolutional Neural Network (CNN) and Bidirectional Long Short-Term Memory (BiLSTM). These models underwent rigorous training and testing to accurately discern individuals based on their handwriting characteristics. Moreover, our methodology introduces a groundbreaking hybrid model that synergizes CNN and BiLSTM, capitalizing on fine motor features for enhanced individual classifications. Crucially, our experimental results underscore the superiority of our approach. The CNN, BiLSTM, and hybrid models exhibited superior performance in individual classification when compared to prevailing stateof-the-art techniques. This validates our method's efficacy and underscores its potential to outperform existing 1690 CMES, 2024, vol.140, no.2 technologies, marking a significant stride forward in the realm of individual identification through handwriting analysis.