Research On Multi-task Face Detection Algorithm Based On Deep Learning In Complex Background

As a key part of face recognition system,face detection has widespread and urgent practical needs,such as identity authentication,security protection,media and entertainment,image search and so on.Driven by the realistic demand,various face detection algorithms have been proposed and achieved good results.However,most face detection algorithms proposed by researchers mostly detect face under strong constraints,face images of these algorithms show a straightforward face with simple background and similar size.However,face images in practical applications face various complicated factors such as light changes,blurring,occlusion,scale changes,atypical gestures,these will influence face detection.Therefore,we need to propose a face detection algorithm that has strong adaptability in complex contexts.In this paper,In order to improve the effect of face detection under complex background,we proposes two efficient multi-task face detection algorithms in view of the characteristics of blur,light,pose,expression and occlusion in complex scene:First,we propose a multi-task face detection algorithm based on Res Net(MRF-CNN).Different from the previous face detection algorithms,the proposed algorithm MRF-CNN has high performance and robustness under complex background such as scale change,illumination change,blurring,occlusion,atypical posture and tiny face etc.MRF-CNN cascaded ML-net and F-net two modules.ML-net improves face two classification and regression tasks by improving Resnet,and improves the accuracy of face detection.F-net can further improve the accuracy of ML-net output results by constructing auxiliary tasks of facial feature points location,which ultimately improves the performance of MRF-CNN algorithm.Second,we presents a face detection algorithm based on super-resolution pyramid(SRPN-CNN)in this paper.During the process of building Pyramid,SRPN-CNN improves the MRF-CNN by super-resolution technology.Our algorithm uses three convolutional networks to reconstruct the low resolution images,then recover the high frequency information lost in the image during the sampling process and provide more detail and making up for some defects in the general interpolation algorithm.The experimental results show that the effects of SRPN-CNN on the change of illumination,blur,occlusion,scale change,atypical attitude and other complex factors than MRF-CNN and other algorithms has better visual effect and predictive performance,which shows that SRPN-CNN is effective for the improvement of MRF-CNN.