Gender classification by deep learning on millions of weakly labelled images

When analysing human activities using data mining or machine learning techniques, it can be useful to infer properties such as the gender or age of the people involved. This paper focuses on the sub-problem of gender recognition, which has been studied extensively in the literature, with two main problems remaining unsolved: how to improve the accuracy on real-world face images, and how to generalise the models to perform well on new datasets. We address these problems by collecting five million weakly labelled face images, and performing three different experiments, investigating: The performance difference between convolutional neural networks (CNNs) of differing depths and a support vector machine approach using local binary pattern features on the same training data, the effect of contextual information on classification accuracy, and the ability of convolutional neural networks and large amounts of training data to generalise to cross-database classification. We report record-breaking results on both the Labeled Faces in the Wild (LFW) dataset, achieving an accuracy of 98.90%, and the Images of Groups (GROUPS) dataset, achieving an accuracy of 91.34% for cross-database gender classification.