Understanding Face Age Estimation: humans and machine

Face age estimation is an important part of many disciplines, including dermatology, cosmetology, and computer vision. Traditional age estimation studies focus on certain parts of the face to analyse its surface topology. With the advances of deep learning, many Convolutional Neural Networks (CNNs) now outperform traditional methods in the age estimation task. However, it is still not clear what type of features these networks learn when estimating age. This study aims to investigate which facial features are important, for humans and for CNNs, on the age estimation of women in five age groups. We then compare the heat-maps from the human eye gaze with those of the CNN. We consider two main research questions: (1) Which facial regions do humans look at when performing age estimation? (2) Do humans and machine focus on the same facial regions when estimating the age? We answered these questions by conducting two experiments. In the first experiment, we used an eye-tracking software to detect where on the face the gaze of human participants focused the most when they were asked to assign the person in the image to an age class. In the second experiment, we used transfer learning on the network pre-trained on ImageNet, then fine-tuned the network on a benchmark face age dataset to classify the same images shown to the participants. The heat-maps of the VGG16 network were then visualised using Gradient-based Class Activation Map (Grad-CAM). The results showed how our model was almost as accurate as humans in predicting the age of a person from a single image of their face (CNN: 60%; humans: 61%). The results also showed that people mainly look at the eyes and nose when predicting a person's age, while the features learned by the CNN included the eyes, the mouth, and the skin surface.