Cascade Regression Face Detection Algorithm Based On Feature Pyramid Network

Face detection is the first step of facial analysis problems such as face recognition and facial reenactment,therefore a precise face detector is prerequisite and important foundation for solving facial analysis problems.This paper proposes cascade regression face detection algorithm based on feature pyramid network,and its main contents are as follow:This paper adopts aggregated residual transformations to build full convolution network as base network.Next,based on the multi-scale and multi-level pyramid hierarchy of deep convolution network,this paper uses lateral connections and upsampling operations to build feature pyramid with high-level context at all scales.Then,this paper exploits position-sensitive operation to introduce translation invariance into full convolution network.Finally,this paper proposes face detection algorithm based on feature pyramid——FPN-RFCN face detection framework.Based on FPN-RFCN face detection framework,in order to enhance context information and improve representation ability of face,this paper combines recurrent neural network with feature pyramid,and designs two kinds of deep recurrent convolution:(1)recurrent up-sampling convolution:it fuses features of adjacent layers by recurrent network,and gradually and selectively introduces deep context information.And comparing with original coarse feature fusion,it can help to improve robustness of algorithm.(2)recurrent rolling convolution:based on recurrent up-sampling convolution,this paper adds down-sampling operations to design recurrent rolling convolution,which introduces fine-grained information to improve representation ability of face.And recurrent structures share parameters with each other,which only brings marginal extra cost.In order to obtain finer regression results,this paper designs structure of three-stage cascade regression,and they share position-sensitive operation without much extra cost and improve localization accuracy of detector.The proposed algorithm achieves average accuracy rate of 99.64%on AFW dataset,and recall rate of 97.08%on FDDB dataset,which proves that the proposed algorithm achieves good performance and is robust to faces with large variations in scales,poses,occlusion,and illumination.