Real-time Face Detection And It's Application

This paper proposes a multi-task Convolutional Neural Network based face detector, which is named "FaceHunter " for simplicity. The main idea is to reach real-time detection while making the face detector achieve a high detection accuracy and obtain much reliable face boxes. To reach this goal, a deep CNN network is designed with a multi-task loss, i.e., one is for discriminating face and non-face, and another is for face box regression. The traditional Convolution Neural Network has very high computational complexity, so it can't meet the real-time requirements. What is more, the above-mentioned two task are learned in two different stage, which increases the difficulty of training. To solve these problems, this paper has made four aspects of improvement:A multi-task loss function is applied to discriminate face/non-face and regress the face bounding box in the same stage sharing the convolution features, which directly simplifies the network's training process; An Adaptive Pooling Layer is added before full connection to make the network more flexible during fine-tune as all layers can be updated in the process of back-propagating. It is adaptive to variable candidate proposals, so the network only needs single time convolution computation; The SVD decomposition is applied to compress the parameters of the fully-connected layers; To further speed up the detector, the convolutional feature map is directly used to generate the candidate proposals by using a Fully Convolutional Network sharing the convolution features with the multi-task Convolutional Neural Network.The proposed detector FaceHunter is evaluated on the three public available face datasets, i.e., AFW dataset, FDDB dataset and Pascal Faces respectively. In order to further validate the practical effect of FaceHunter in the auto face analysis system, it is also applied in the real-time video driven 3D facial animation system with the face alignment technology. The extensive experiments and the practical effect demonstrate its powerful performance against several state-of-the-art detectors.