Development Of Convolutional Neural Network-Based Pelican Optimization Algorithm for Handwriting Identification System

Abstract

Handwriting identification remains a significant challenge in the field of information processing and Optical Character Recognition (OCR) due to the diverse nature of human writing styles. Traditional Convolutional Neural Network (CNN) models, although widely used, often suffer from overfitting, hyperparameter sensitivity, and limited adaptability to different handwriting patterns. This research addressed these challenges by optimizing Convolutional Neural Network (CNN) with Pelican Optimization Algorithm (POA) for Handwriting identification System (HRS). A handwriting dataset comprising 3000 forged handwriting samples and 3000 original handwriting samples was obtained from Kaggle,com. The images were resized, normalized, and augmented to enhance model generalization. POA-CNN was developed by using Pelican Optimization Algorithm to fine-tune CNN hyperparameters of learning rate, layer size, activation functions, and batch sizes with Keras support packages. The POA-CNN model was implemented in MATLAB R(2023a).The system’s performance was evaluated across varying decision thresholds with six key metrics employed: False Positive Rate (FPR), Specificity (SPEC), Sensitivity (SEN), Precision (PREC), Accuracy (ACC), and Computational Time (CT). The model was compared with traditional CNN. The optimum threshold was 0.51. The FPR, SPEC, SEN, PREC, ACC and CT for POA-CNN were 3.20%, 96.80%, 96.70%, 96.80%, 96.75%, and 81.95 s, respectively. The corresponding values for CNN were 4.77%, 95.23%, 95.13%, 95.23%, 95.18%, and 91.20 s, respectively. The developed POA-CNN for handwriting identification system demonstrated better performance than the traditional CNN, across all metrics. Overall, the results demonstrate that POA-based hyperparameter optimization significantly improves the accuracy and reliability of CNN-based handwriting identification for effective forgery detection.