Learning Based Age Estimation Using Joint Loss and Facial Landmarks

Age recognition is an important technology in computer vision, surveillance systems, and commerce. It can be applied in many scenarios, including age restrictions in specific places, drinking restrictions, and reminders for underage or elderly drivers in traffic applications. In this study, we proposed an accurate age recognition algorithm, which applies the techniques of landmark-based alignment, the attention model, and the expected value method in the deep learning architecture. The algorithm consists of three stages. The first stage is data preprocessing, including face detection, cropping, face alignment (to normalize the position and angle of each face), and contrast adjustment. The second stage is a feature extraction model. It is based on the Residual Attention Model with the attention mechanism. Moreover, the domain filter, the joint loss, and the facial landmark extraction are also adopted. The third stage is the classification model. Its input is the l024-dimensional features extracted in the 2 ^nd stage and the input image. Then, the expected value method is applied to calculate the explicit age. Simulations show that the proposed algorithm outperforms other age estimation methods and can estimate the age accurately.