Multi-view Cooperative Deep Convolutional Network for Facial Recognition with Small Samples Learning

In the community of computer vision, deep learning has been widely applied in many classification tasks. However, the performance of deep networks depends heavily on the large number of labeled samples. In this paper, we propose a multi-view Deep Convolutional Neural Network to recognize facial expression while very small number of samples is available. First, facial images are downsampled to different scales and upsampled as multi-view samples. Then a multi-view DCNN is constructed with twin structure and cooperative learning. After one channel is trained by single view samples, the parameter is transferred to another channel for fine tuning using another view samples. Some experiments are taken on FER2013 and RAF datasets, and the experimental results illustrate that the proposed multi-view DCNN network has a good performance where achieves 72.27% on the private set of FER2013 dataset, and the transfer DCNN model achieves 83.08% on the test set of RAF database.