Experimental Recognition Time Analytical Evaluation of CO with CNN on Handwritten Signature Validation Performance

Abstract

The issue of identity recognition using handwritten signature biometrics behavioral features has been an important aspect in day-to-day human activities. However, disparities in signature recognition and identification sprouts out greatest challenge due to health instability status in human and unpredictable imitations initiated by forgers. Convolutional Neural Network (CNN) however as one of the prominent deep learning architectures in image processing tasks was applied in this research to the acquired 5220 signature datasets; comprising 2750 Genuine signatures and 2470 Forged signatures. Inadequacies identified in CNN model; which includes overfitting, learning rate and recognition accuracy is then addressed by optimizing the CNN parameters using Cheetah Optimization Algorithm (COA). The image datasets were simulated with MATLAB R2023a and evaluated by adopting Confusion Matrix six evaluation metrics - False Positive Rate (FPR), Precision, Sensitivity, Specificity, overall Accuracy and Recognition Time. Affirming the resulting correlations of the two models, mathematical equations were derived using curve-fittings tools in MS Excel, based on derivable values from CNN and CO-CNN using recognition time in other to establish the system accuracy. Therefore, for recognition time analysis, the complexity associated with CO-CNN is less than that of CNN thereby improving the balance between the exploration and exploitation stage of the technique. In general, the overall results indicate that the CO-CNN performs more better than CNN in terms of Precision, Specificity, Accuracy and recognition Time performance metrics.