Research On Face Detection Based On Deep Learning And Feature Fusion

Face detection is a key technology in face information processing,it is a prerequisite and foundation for many face image analysis applications such as face recognition,face alignment,face tracking and face attribute recognition(gender,age,expression),and it’s also the crosscutting research topic of applied mathematics and computer vision.The paper proposes a face detection method based on deep learning and feature fusion.Firstly,we propose a deep multi-layers feature fusion convolutional neural network based on the deep convolutional neural network(DCNN).In this network,the structure of Inception and Skip connection is adopted to increase the depth of the network and reduce the model parameters,which alleviates the gradient vanishing problem.And we use the feature pyramid network to combine the high-level semantic information and the low-level detail information to enrich the feature representation.Experimental results show that the design and composition of the network structure outperform the other basic feature extraction network.Then,based on the Faster RCNN detection framework,we use the above mentioned deep multi-layers feature fusion convolutional neural network as the basic network.In the candidate region extraction stage,the multi-region feature fusion method is used to capture different aspects of the face to obtain a abundant representation and named Multi-region Proposal Network(MRPN).It’s helpful to improve the face detection performance,and combined with online hard example mining(OHEM)to further improve the quality of candidate regions.Furthermore,we design the Block Loss for different candidate regions to enhance the attention of the different candidate regions,so that the model can be robust to the complex situations such as occlusion.And we use the Soft-NMS method for the post-processing of the detection results at last.The experimental results show that the Block Loss and Soft-NMS method can effectively improve the final detection results.Finally,the experiments prove that our method has better performance for face detection with small size,severe occlusion and blur.In the FDDB face detection standard database,the recall rate reached 96.10%when the number of false positive is 500,and the average accuracy rate achieved 99.48%on AFW.