A Compact Soft Ordinal Regression Network for Age Estimation

Hang-Xing Zang, Han Su, Yu Qi, Hai-Kun Wang

Abstract: Age estimation is a focal point of computer vision research. With the development of Convolutional Neural Networks (CNNs), some previous studies utilize large-scale models to learn the informative age representations to boost performance. However, these models with a large number of parameters are not suited to the low-power devices. In this paper, we propose a compact model called Soft-Lite ECA RepVGG network (LERep) for age estimation to alleviate this problem. We utilize the re-parameterization technique and the Efficient Channel Attention (ECA) module to construct a novel model. Decoupling the training and inference stage can reduce the computation cost. In other words, the training-time model is a multi-branch model, and then we convert the multi-branch model to a single plain model during inference time. Our approach has been proven to effectively trade off the performance and parameters for age estimation. Moreover, we propose a novel loss that utilizes the ordinal age information and expectation of age label distribution to address the non-stationary aging problem. According to the experimental results on MORPH II, FG-NET and CLAP2015, our models achieve state-of-the-art performance compare with other compact models. In particular, our models also achieve competitive performance compare with over larger models, while our models have approximately 1/250 parameters of the VGG-16 network.