Transfer Learning with EfficientNet for Computer Vision based Automatic Gender Classification

The past several years have seen a lot of activity in the field of gender classification research. For a wide range of tasks including surveillance, computer vision, human-computer interaction, and traffic monitoring, gender classification is essential. The traditional approach used to classify gender has a number of drawbacks and is insufficient when used with live images. However, gender classification is just one of several computer vision applications where deep learning algorithms have demonstrated their potential. CNNs are capable of automatic feature generation and provide data insight which is used for efficient classification. In this paper, a special class of pre-trained CNN, called EfficientNet is proposed for automatic gender classification from input image. The experiment is carried out using a dataset with 58,700 images categorized into male and female. The performance of EfficientNet-B0 is also compared with other benchmark pre-trained CNN models i.e. VGG16, ResNet50 and DenseNet. The Precision, Recall and F1-Score values achieved for EfficientNet-BO are 96.80%, 96.60% and 96.88% respectively.