Multi-Scale Similarity Learning for Age Estimation Based on Facial Images

The facial age estimation is an important task for efficient human-robot interaction. A supervised similarity learning has not been widely adopted to a facial age estimation and is still a challenging task. In this paper, we propose a multi-scale similarity learning model which is able to jointly learn similarities between multi-level age features extracted from three different images: an input, a sample with same age named positive and a sample with contrary age named negative. The idea behind using the multi-level feature extraction is to make our feature similarity learning model robust to scale invariance of individual images. As is known, the quality of images in datasets for age estimation vary considerably and deep-learning based networks are often sensitive to image quality and resolution. On that account, we employ a multi-scale feature extraction structure to our model and prove its ability to extract reliable features for a similarity learning. Experimental results show that the proposed approach outperforms previous research and provides a new state-of-the-art age estimation accuracy of UTKFace and CACD benchmark datasets for age estimation.