Comparative evaluation of AlexNet, SqueezeNet, VGG16, and ResNet50 for gender and hijab detection

Abstract

This study aims to detect gender based on facial images with and without hijab features, with the expected outcome of distinguishing gender from these facial features. The method involves comparing the performance of four convolutional neural network (CNN) architectures: AlexNet, SqueezeNet, VGG16, and ResNet50. A total of 170 facial images were directly collected using smartphone cameras. The dataset consists of two classes: 68 male faces and 102 female faces, among which 78 images of females feature hijabs, while 24 do not. The validation stage with 40 images (15 males and 25 females) showed that the AlexNet architecture achieved the highest validation accuracy at 100%, followed by ResNet50 with 97.50%, VGG16 with 95%, and SqueezeNet with 92.50%. The testing stage with 40 images (20 males and 20 females, including 10 females with hijabs and 10 without) showed that ResNet50 classified 38 images correctly, achieving 95% accuracy. AlexNet classified 37 images correctly with 92.50% accuracy, SqueezeNet classified 36 images correctly with 90% accuracy, and VGG16 classified 34 images correctly with 85% accuracy. The contribution of this research shows that AlexNet achieves the highest validation accuracy, while ResNet50 provides the best accuracy in facial image detection for determining gender and hijab features.