Research On The Algorithms Of Face Quantitative Detection

At present,the face detection algorithm is to qualitatively determine whether there is a face in a single image or a video frame.With the expansion of the application scope of face detection,only detecting the presence of face in the image does not satisfy the need for precise positioning of some image editing and video editing applications,and more importantly,we need to improve the detection results in the quantitative description of the specific positions and angles of the faces.In the actual video face replacement project,the traditional face detection algorithm ignores the time consistency of face location and angle in video,causing the flicker of replacement faces.Based on the above research,this topic starts from the perspective of face quantitative detection in video,and quantitatively detects the position and angle of the face in the source video and the target video respectively,so as to obtain the specific position and deflection angle of the face.The main works of this paper are as follows:In the quantitative detection part of the face position,a quantitative detection algorithm for video face location based on inter-frame constraint model is proposed.This paper firstly uses the Adaboost algorithm based on Haar?like feature to detect the face in video,since the result of face detection between two adjacent frames in video sequence is usually biased,the model considers the correlation of face position between frames in the video,constraint face position of the detected face,so as to give quantitatively the specific face position information to avoid the deviation of single frame detection and ensure the time consistency of the face position in the video.An adaptive face search region model is proposed to improve the speed of video face detection.The detected face position and the vicinity of the previous frame are used as candidate regions for face detection in the next frame,so as to adaptively change the face search area and avoid the invalid non-face range.Compared to traversing the entire frame image to detect face the time efficiency of this algorithm is higher,to a certain extent,it meets the requirements of real-time video face detection.Two kinds of Haar?like EB(Haar?like Eye and Brow)features are proposed,which effectively solved the false detection of face detection algorithm based on Haar?like features due to the similarity of eye and eyebrow gray value.Experiments show that the algorithm improves the performance of the face classifier and reduces the false detection rate.In this paper,convolution neural network is used as the main method to design a model for quantitative detection of face angle.By referring to the classical VGG convolution neural network model and combining with the characteristics of face images in the direction of composite angles,the network hierarchy is improved,and a 22-layers depth convolution neural network model,Face Angle Net,is proposed.The model can classify the face angles in the video in both pitch and yaw directions,so as to realize the quantitative detection of face angle in the video.By using CAS-PEAL-R1 face database for training,the classification accuracy rate on the verification set is 96.45%,which is higher than other algorithms.