Research On Face Detection From The Front-Top Perspective Of UAV

Unmanned aerial vehicles(UAV)needs to evaluate and track the face of the target when doing security detection,so detecting faces fast and accurately from the front-top view of UAV is the basis of related face tasks.Starting from solving problems of face detection from the front-top view of UAV,this paper carried out a series of related works.Firstly,this paper set a face detection dataset from the front-top view of UAV,compared the results of several leading methods of face detection on this dataset,and analyzed the characteristics and difficulties of face detection from the front-top view of UAV.Based on the characteristic that face detection from the front-top view of UAV has more small faces and high ratio of background,this paper presents an improved cascaded face detection algorithm.We combined single-shot proposal net with multi-layer information,which can propose different scales of regions on different layers of the net;We also improved the strategy of deleting redundant overlapped face regions,which increased the ability of detecting highly overlapped faces of the net;And we used more effective classification loss function to classify faces during the face detection process,which increased the accuracy of face detection from the front-top view of UAV effectively.Next,because faces from the front-top view of UAV has abundant contextual information of human body,this paper added contextual information of human body to the region proposal net and the refine-and-output net of the cascaded face detection frame.Moreover,attention mechanism is added to all the stages of the cascaded face detection algorithm as supervised signal,which can highlight the features which have important effects to the face detection from the front-top view of UAV.The combination of the above two methods enormously increased the accuracy of the cascaded face detection algorithm from the front-top view of UAV.Finally,Generative Adversarial Nets network is added to the face detection from the front-top view of UAV.Based on the proven fact that the contextual information of human body has positive effects to the face detection results from the front-top view of UAV,the samples of small faces with contextual information of human bodies generated by generation antagonistic network is used to train the cascaded face detector,which improved the ability of the net detecting small faces effectively.What’s more,generation antagonistic network is also used to do super-resolution reconstruct of potential small face samples.The reconstructed photos are sent to the cascaded network for further detection,which improved the accuracy of the cascaded face detection algorithm detecting small faces effectively.