Deeper cascaded peak-piloted network for weak expression recognition

Facial expression recognition is in general a challenging problem, especially in the presence of weak expression. Most recently, deep neural networks have been emerging as a powerful tool for expression recognition. However, due to the lack of training samples, existing deep network-based methods cannot fully capture the critical and subtle details of weak expression, resulting in unsatisfactory results. In this paper, we propose Deeper Cascaded Peak-piloted Network (DCPN) for weak expression recognition. The technique of DCPN has three main aspects: (1) Peak-piloted feature transformation, which utilizes the peak expression (easy samples) to supervise the non-peak expression (hard samples) of the same type and subject; (2) the back-propagation algorithm is specially designed such that the intermediate-layer feature maps of non-peak expression are close to those of the corresponding peak expression; and (3) an novel integration training method, cascaded fine-tune, is proposed to prevent the network from overfitting. Experimental results on two popular facial expression databases, CK+ and Oulu-CASIA, show the superiority of the proposed DCPN over state-of-the-art methods.