Face Detection And Recognition For Video Images

In the present,the research on face images has advanced by leaps and bounds,and with the development of artificial intelligence,related technologies in this field have been applied to the public's life and tend to mature.This thesis studies the face detection and recognition technology for surveillance video.At the same time,in order to improve the accuracy and speed of recognition,an image screening algorithm is proposed to be added to the framework of face detection and recognition,and an improved kernel is proposed for the characteristics of target loss in the process of face detection and tracking.Correlation filter tracking algorithm.The face detection recognition and tracking system implemented in this thesis firstly performs face detection on the input video image.Secondly,in the process of detection,the best positive face extraction algorithm is used to screen each detected face and select one.The most suitable face image is input into the recognition network,and the identity of each object is determined and a tag is generated on the face frame in the video.At the same time,the result of the detection is passed to the tracker to achieve continuous tracking of the target object.In addition,in order to meet the situation that the monitoring video often enters and leaves the frame,the above detection and tracking process is repeated in a periodic form,and when a new "unknown face" appears,the identification module is called again to identify the identity.The main work of this paper includes:(1)In order to adapt to the application environment of face detection for video,the weight self-adjustment algorithm is continuously adjusted in the process of implementing multi-task convolutional neural network(MTCNN)face detector.The key parameters in the MTCNN network get the most suitable parameter settings.Through testing,it is proved that the MTCNN model implemented in this thesis meets the expectation of high efficiency and high accuracy for face detection in video.(2)In order to improve the efficiency and accuracy of face recognition,a face detection and recognition framework is established.An image screening module is added between face detection and face recognition.In order to avoid the problem of the recognition effect caused by the angle and quality of the face image,this thesis gives the best positive face by calculating the positional relationship of the face feature points,the image clarity and the probability that the framed image is the face.Extraction algorithm.The algorithm is used to filter the pictures of the input recognition network,thereby improving the performance of the overall recognition process.(3)An improved kernel correlation filtering algorithm is proposed for the problem of easy target loss in face detection and tracking.First,use the peak sidelobe ratio to calculate whether the target is lost.Secondly,if the tracking loss is judged,it is resolved in two cases: the first type,partial occlusion: re-detection by the MTCNN detector;second,complete occlusion: the image is pre-processed by genetic algorithm,and then The Kalman filter is used to predict the current position of the target based on the state information of the target in the adjacent video frame.