A Single Shot Multi-Head Gender, Age, and Landmarks Detection using Shared Convolution Features

Considering the face as a vital and most informative portion of the human body, it reflects different high-level information about an individual. This high-level information includes Age, Gender, and Emotion. Facial muscles’ shape and movement can be the best descriptors for the automatic extraction of these high-level facial features. Detection of these high-level features has applications in different areas including entertainment, surveillance, multimedia, and educational training. However, with the varying nature of these features, it becomes difficult to capture one class with the variability of other classes. This article presents a lightweight heterogeneous neural network with one shared backbone and three network heads to predict multiple face features: landmarks, age, and gender. The proposed system (MultiHeadCNN) captures these high-level facial features in the wild with extreme face pose, occlusions, and lightening conditions. The system is capable of predicting one type of feature with different variability of other types: predicting gender for different age groups and vice versa. The system is tested on comprehensive (UTKFace) and complex (Adience) datasets with varying age, gender, pose, and lightening conditions. The experiment shows promising results in terms of accuracy, with results for age and gender detection on the UTKFace and Adience datasets being 99.9%, 99.7%, 90.3%, and 61.7%, respectively. Furthermore, the parallel inference speed is 20 frames per second.