AE-Conv MLP: A Lightweight Convolutional MLP for Age Estimation

Age estimation from facial images is an attractive research topic in computer vision. A variety of large and deep Convolutional Neural Networks (CNN) based approaches have been proposed with promising performance. However, these models are not suitable for deploying on mobile platforms with limited resources. In response to this weakness, we propose AE-ConvMLP, a lightweight Convolutional Multi-Layer Perceptions (MLP) architecture for Age Estimation, which can be effectively applied to mobile and embedded devices. Inspired by the long-range modeling ability of MLP, AE-ConvMLP integrates the strengths of CNN and vision MLP to capture both short-range and long-range visual dependencies. Concretely, the depthwise convolution is employed to capture essential local visual cues, and a novel Fully Connected (FC) layer named Global Region FC (GRFC) is proposed to efficiently aggregate global visual dependencies by a sparse token partition strategy. To achieve a stable learning, we propose a new loss named Gaussian Error loss, in which the difference between the predicted and the ground-truth value is regarded as a distribution rather than a discrete value to enhance stability at the training stage. Finally, we optimize our network in a joint learning fashion by performing regression and label distribution learning simultaneously. Extensive experiments on several commonly-evaluated age estimation datasets are performed. Our AE-ConvMLP achieves state-of-the-art performance among small models with lower complexity and outperforms some large models by a relatively large margin with only 1/300 parameters and 1/1519 FLOPs.