| Face detection plays a decisive role in the entire field of public security.With the development of computing power and annotation datasets,many researches have been made to develop face detection algorithm based on deep learning.However,in practical application scenarios,face detection encounters problems such as multi-scale targets,small-size targets,and dense targets.In addition,the face detection algorithm based on deep learning requires a large amount of computing resources,while the terminal application scenarios of security,such as street monitoring and dash cameras,often have limited resources.In order to solve the above problems,this thesis deeply studies two aspects---the face detection algorithm based on convolutional neural network and the compression algorithm of convolution neural network.In the face detection algorithm,the algorithm proposed in this thesis,Zface Pro,has the following four features:1.Propose a detection framework with the face prior knowledge.Based on the object detection algorithm SSD,this paper proposes a face detection framework with the prior knowledge of face shape and distribution.2.Use random scaling to augmentate the dataset online to improve the ability of the algorithm to recognize multi-scale targets.3.The soften non-maximum suppression algorithm is applied to the post-process of the detection and enhances the ability of the detection algorithm to detect dense objects.4.Optimize algorithm engineering implementation,reduce memory consumption and speed up execution.The face detection algorithm Zface Pro proposed in this thesis has achieved a precision of 94.5% with 100 false positives on the public test set FDDB and an execution speed of over 70 FPS on the Titan X.As for network compression,this thesis has made the following two contributions:1.In a single layer,only xor binary convolution is used instead of floating-point convolution.With the help of CPU hardware instruction popcnt,this thesis’ s algorithm can execute as speed as open-source floating-point computing library Open BLAS.2.Use mixed precision in the complete network.The combination of floating-point convolution and xor binary convolution not only reduces the memory consumption of the network,but also retains the generalization ability of the network.Zface XOR,which is based on the mixed-precision compression scheme proposed in this paper,achieves a precision of 94% with 100 false positives on the FDDB test. |
