Deep Learning Architecture of VGG16 and VGG19 for Eyeglasses Face Classification

Abstract

This study aims to conduct a comparative analysis of the performance of two popular Convolutional Neural Network (CNN) architectures, namely Visual Geometry Group (VGG16 and VGG19), in classifying facial images with glasses using the “Glasses or No Glasses” dataset. Both models were developed through a transfer learning approach by utilizing pre-trained ImageNet weights to accelerate convergence and improve classification accuracy. The training process employed the Adam optimizer with binary crossentropy as the loss function. The dataset was divided into two subsets 80% for training and 20% for validation while testing was performed on 50 unseen images excluded from both subsets. Experimental results show that the VGG16 architecture achieved 87.86% training accuracy and 89.11% validation accuracy, whereas VGG19 achieved 86.86% training accuracy and 87.89% validation accuracy. On the testing dataset, VGG16 correctly classified 47 out of 50 images (94%), while VGG19 correctly classified 48 images (96%). Although the performance gap is relatively small, VGG19 demonstrated better computational efficiency with a shorter training duration (2 hours and 41 minutes) compared to VGG16 (2 hours and 59 minutes). Furthermore, the trained models were successfully implemented in an Android application using TensorFlow Lite, enabling real-time eyeglasses detection. These findings indicate that the VGG19 architecture offers superior efficiency and accuracy for deep learning–based eyeglass face classification tasks.