Detecting Minors According to South African Law Using Computer Vision Methods

Age estimation is one of the areas of interest in computer vision, which is evident from the increased amount of related research over the last few years. This is largely due to the exceptional levels of classification accuracy demonstrated by Convolutional Neural Networks (CNNs) in computer vision tasks. One of the main challenges faced when training age detection models are accounting for people’s diversity, which raises the importance of using datasets with as much diversity as possible. Another important factor to consider is the reason behind performing age estimation, which can either classify people’s age or detect if someone’s age exceeds (or is below) a specific threshold. This paper presents work done to detect minors according to South African law (which is under 18 years of age). South Africa is a very diverse country. As such, the UTKFace dataset, containing face images with a wide range of ethnicities, was used to train the Inception Resnet V2 to detect minors according to South African Law. The dataset was reduced to only include relevant images with the aim of obtaining an equal distribution of gender and ethnicity to ensure relevance to the South African context. A model accuracy of 99.73% was achieved, demonstrating the model’s ability to distinguish between underage and legal age classes. It was also noted that class imbalance and the reduced number of samples were inhibiting factors to the model’s performance in terms of precision and recall.