Noisy facial recognition with convolutional neural network (CNN)

• Main Author: Heng, Yin Qi
• Other Authors: Anamitra Makur
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Engineering; Machine learning; CNN; Facial recognition
• Online Access: https://hdl.handle.net/10356/176638
• Institution: Nanyang Technological University

The aim of this research is to evaluate how Gaussian noise affects the accuracy of two Convolutional Neural Network (CNN) architectures—LeNet and AlexNet—in facial recognition applications. By utilizing the Faces95 dataset, the networks were subjected to varying levels of Gaussian noise to determine how noise impacts model performance. The study revealed that while LeNet was moderately resilient to noise, AlexNet's accuracy suffered significantly. Specifically, at a noise variance of 0.05, AlexNet experienced a marked accuracy reduction compared to LeNet. This divergence highlights the importance of architectural complexity in noise sensitivity. Importantly, the study shows that incorporating noise during training can significantly improve the noise robustness of both models, indicating that noise-adaptive training is a powerful strategy for enhancing CNN performance in visually noisy environments. These findings deepen our understanding of CNN resilience against visual noise and underscore the potential of revisiting simpler CNN architectures to unlock and leverage innate noise-resistant features. The insights gained here pave the way for developing more reliable facial recognition systems that can function effectively in real-world noise disturbances.