Statistical Comparison of Manual and Machine Learning Signature Verification Using Neural Networks on Raspberry Pi

This study presents a statistical comparison between a machine learning-based handwritten signature verification system, and a traditional manual verification method. The research aims to compare and evaluate signatures by applying machine learning and image processing techniques. It integrates a Raspberry Pi 3 Model B+ with a camera module and a touchscreen interface to capture and process handwritten signatures. Three neural network algorithms—LeNet, Feed Forward, and VGG16 Transfer Learning—were trained and tested using a dataset of 4,668 genuine and forged signature samples. Image pre-processing techniques such as grayscale conversion, thresholding, and noise reduction were employed to enhance data quality. Among the three algorithms, the LeNet model achieved the highest accuracy of 99.36%, outperforming the Feed Forward (99.20%) and VGG16 Transfer Learning (99.14%) models. Using the LeNet model, a comparative assessment was conducted between the manual verification method and the machine learning-based verification system. The Mann-Whitney U test confirmed that the verification system showed significantly higher accuracy and reliability in verifying genuine and forged signatures. The null hypothesis—that there is no difference between the manual and system methods—was rejected at a significance level of 0.05. Furthermore, system acceptability was assessed through a survey conducted among technical experts, focusing on usability, convenience, interface design, and output accuracy. The survey demonstrated acceptable reliability (Cronbach’s alpha = 0.728) and produced a grand mean score of 4.6 out of 5, indicating a “Very Good” perception of the system’s performance and user-friendliness.