Face Detection Using Multi-superised Information And Cascade Fully Convolutional Networks

Face detection is one of the key research fields in the computer vision research.It has great research value in both scientific research fields and the commercial fields.In scientific research,detection problem is the fundamental problem to the other computer vision problems.Face detection,which is treated as the fundamental problem of face alignment,and face verification problem,has attracted great attention.In the commercial fields,faces are expected can be detected and recognized automatically and accurately in many different scenes such as security,finance,bayonet etc.However due to the complexity of the real scene,and the uncertainties of face pose,angle,position,background,there are still many difficulties for face detection not to be solved completely.Face detection problem can be divided into two part,how to train a binary face classification and how to find face in the image.The time of face detection algorithm can be seen as the superposition of two sub-problem.At present,face detection algorithms are in a dilemma.Some of their performance are good while their test time are long because the model is too complex;while others can reach real-time,but the performance are not good because of their relatively simple model.It is significant to make a fast and good face detection algorithm.The current main method treat the two sub-problems independently,that is the binary classification model is based on the deep convolutional network,and the search strategies of face are using sliding window,selective search and so on,which make the current methods has many limitations like excessive consumption of computing resources or long time for detecting.In this paper,we propose a face detection algorithm that is simple and effective and can detect multiple scales simultaneously.We try to combine the two subproblems together and train an end-to-end face detection model by using FCN without restricting the scales of faces.Compared with general convolution neural networks,the fully convolutional network has the advantage of spatial position information.Therefore,fully convolution network is more suitable for tasks such as detection and segmentation.In this way,there is no need to rescale image in multi sizes in the test phase.In addition,in order to make the model better training,we try to divide face detection problem into two sub-problems.Namely face classification sub-problem and face bounding box regression sub-problem.We propose multi-view supervised information to help the network training,and finally make a cascade fully convolutional face detection algorithm based on multi-view supervised information.Extensive experiments suggest the effectiveness of proposed methods.