Learning Saliency Features for Face Detection and Recognition Using Multi-task Network

Abstract

In this work, we have proposed a method to learn a type of saliency features, which merely makes response in face regions. Based on the saliency features, a joint pipeline is designed to detect and recognize faces as a part of human–robot interaction (HRI) system of SRU robot. The characteristics of the architecture can be described as follows: (i) In the network, detectors can only be activated by face regions. By convoluting the input image, the detectors can produce a group of saliency feature maps, which indicate the location of faces. (ii) The face representations are achieved by pooling on these high response regions. They enjoy discriminative ability to face identification. Hence, classification and detection can be blended using a single network. (iii) To enhance the saliency of features, false responses are suppressed by introducing a saliency term in loss function, which forces the feature detector to ignore non-face inputs. It also can be seen as a branch of multi-task network to learn background. By restricting false responses, the performance of face verification can be improved, especially when the training and testing are implemented on different dataset. In experiments, the effects of saliency term on face verification and benchmark discriminative ability of saliency features on LFW are analyzed. And the effectiveness of this method in face detection is verified by the experimental results on FDDB.