| Face detection has always been an important component of computer vision.With the development of artificial intelligence,people are increasingly demanding for face detection algorithms,especially in mobile applications.However,the current face detection algorithms under the mobile terminals have some problems,such as low accuracy and slow running speed,which will affect the implementation of follow-up algorithms or operations.In order to improve the accuracy and the speed of face detection under mobile terminals,a lightweight and fast face detection network is proposed in this paper,which combines the requirements of algorithms research and application environment.The research work and results of this paper are as follows:(1)Firstly,this paper analyses the methods of face detection,and compares each algorithm in mobile terminals.The results show that MTCNN has the best performance under mobile terminals.So we proposal using cascade multi-task structure to create the network.(2)Analyzed and realized the speed of MTCNN algorithm is the biggest problem,and the most time-costing operation is to extract the prediction boxes.It is found that RPN in Fast R-CNN does not increase extra time.Then a network based on pyramid structure is set as stage one of the cascaded network.For the performance of shallow network is not as good as deep network,BN layer is introduced in training,and at the end of training merged with its front convolution layer,thus improving the performance of network without increasing the parameters.(3)Because of the difficulty of training and the poor performance of small networks,a method of combining multi-layers and knowledge distilling is proposed.Singular value decomposition is used to compress the network model to adapt to the limited memory environment of mobile terminals.Finally,the trained model tested on Android platform and achieved 80 fps.Compared with MTCNN and other algorithms,it has advantages in accuracy and speed,and the size of the model is the smallest. |
PDF Full Text Request