Design And Implementation Of Face Detection System For Monitoring Video Based On Normalized Pixel Difference And Deep Learning

With the continuous development of artificial intelligence technologies,various new intelligent devices have emerged,which have attracted wide attention for their convenient and novel user experience.And the application of computer vision technology enhances the functions of the monitoring system,and effectively reduce the manual workload.Based on this,this paper studies the face detection problem under the monitoring scene,and designs a fast face detection system which is applied to the monitoring scene.The monitoring scene is a typical unconstrained scene,the background content of the monitoring video is varied and complicated,the angle of faces and the facial expression are uncertain,the illumination conditions are unstable,and the face image size is relatively smaller.In addition,there are some requirements on the detection speed of the algorithm when applied to the videos.By making a targeted improvement to the existing algorithm,this paper achieves rapid and accurate face detection under different lighting conditions and a certain range of angles in the monitoring scene.Firstly,this paper presents a facial feature based on multi-scale joint NPD feature,which reserves the advantages of NPD features such as low computational complexity and quick solution speed,and can reduce the impact of face detail changes on the detection effect through multi-scale changes,so as to complete the face detection function better under non-constrained monitoring scenarios.Secondly,this paper designs and implements a face detection network based on the Faster RCNN framework,which enhances the detection speed of the original Faster RCNN,and verifies the performance of face detection algorithm based on the deep neural network in real-time videos.Besides,in order to speed up the processing speed of the system,this paper designs and realizes a framework of face detection system in multi-threaded environment,which can improve the detection speed of the existing algorithms by detecting the image pyramid of a single image in parallel.In addition,in order to train face models based on regional NPD features and convolutional neural network,this paper constructs a dataset including 170,000 positive samples with human faces under unconstrained scenes and 11404 negative samples without human faces.In this paper,we use this data dataset to train face models based on multi-scale and NPD feature,and test it on FDDB,which has a good detection effect.