Research Of Face Detection Methods Based On Deep Learning

Currently,face detection methods based on convolutional neural networks have achieved great success.However,face detection is still a challenging problem because of faces in images and videos having high degree of variability in scale,posture,expression,occlusion,blur and so on.In order to improve the detection effect of face detection tasks,this thesis:Propose a Hybrid Feature Pyramid.Hybrid Feature Pyramid uses different ways to generate high-level semantic information and feature maps for detection.Compared with FPN(Feature Pyramid Network),Hybrid Feature Pyramid processing features is more detailed.Design a new face detection method called HPCNet.HPCNet introduces a Hybrid Dilated Convolution,a Hybrid Feature Pyramid,and Context Information Extractors into convolutional networks.Hybrid Dilated Convolution can quickly increase the receptive field and obtain high-resolution feature maps.Contextual information helps to improve classification accuracy.Train HPCNet with the improved OHEM(Online Hard Example Mining).The improved OHEM can select positive and negative samples more evenly.In this thesis,HPCNet is evaluated on the WIDER FACE validation set.The average precision(AP)on the Easy,Medium,and Hard subsets is 0.933,0.924,and 0.848,respectively.Face detection tasks not only require the detection effect to be good enough,but also require that the reasoning time is short enough and the detection speed is fast enough,which is another important challenge for face detection.In order to improve the detection speed of face detection tasks,this thesis:Design a Two Information Flow Block(TIFB)and set up a Feature Maps Fast Shrink Module(FMFSM)based on TIFB.TIFB can obtain more robust feature maps and helps to enhance the transfer and reuse of feature maps and the back propagation of gradients.FMFSM slows down the number of feature map channels and quickly reduces the size of feature maps,which greatly reduces the amount of computation on networks.Design a Retinal Receptive Field Block(RRFB)and establish a Variable Scale Face Detection Module(VSFDM)based on RRFB.RRFB mimics the human visual system and can acquire robust feature maps with rich scale information,quickly improve the computational efficiency of networks.VSFDM performs face detection on multiple feature maps,alleviating the burden of the single or composite feature map.Design a new face detection method,named LRNet.LRNet consists of FMFSM and VSFDM.LRNet has a small amount of calculation and high computational efficiency.Improve the priori boxes strategy of Face Boxes.The improved priori boxes strategy considers not only the density of priori boxes of different scales,but also the nature of the features representing the regions.In this thesis,LRNet is tested on FDDB.When the number of false positive(FP)examples reaches 2000,the true positive rate(TPR)at discrete and continuous scores is0.951 and 0.725,respectively.LRNet can run at 112FPS(Frames Per Second)on NVIDIA 1080 TI for 1024*1024 resolution images.