Face Detection And Facial Landmark Localization Based On Improved MTCNN Model

With the rapid development of computer vision,face recognition technology has gradually been accepted by the public and has become a research hotspot of artificial intelligence in recent years.The face recognition system refers to an artificial intelligence technology that analyzes the visual feature information of the face and performs the identification of the person.The technology can be used in many fields,such as public security detection systems,access control systems,work punching systems and video surveillance systems.The face recognition system usually consists of five parts: image acquisition,image preprocessing,face detection,facial landmark localization and face recognition.Because face detection and facial landmark localization of face are two important parts of face recognition system,the performance of these two parts directly affects the accuracy and timeliness of subsequent face recognition.Therefore,this paper focus on the research and discussion of face detection and facial landmark localization.The so-called face detection refers to determining whether there is a face in the input image according to the algorithm or model.If the image contains faces,the areas where faces are located are framed.This paper selects the left and right eyes,the nose tip and the left and right corners as five landmark points.Face detection and facial landmark localization can crop the detected face area and scale the size applicable to the face recognition system,laying the foundation for the subsequent face recognition.Due to the existence of many unconstrained environmental factors such as light,obstruction,non-positive face,and low image pixels,it brings certain difficulties and challenges to the accuracy of face detection and facial landmark localization.In order to enhance the accuracy and robustness of face detection and facial landmark localization,so that the follow-up process of face recognition can make a correct decision.Based on the MTCNN model,this thesis conducts in-depth research on those two tasks.The main research content of this paper consists of two parts: the first part is the research of facial landmark localization method based on brain parallel interaction mechanism;the second part is the new face detection system which combines MTCNN model and ERT model.The main innovations of the paper are as follows:1)Aiming at the single template multi-scale image pyramid preprocessing in the MTCNN model,multi-template multi-scale data preprocessing is introduced into the MTCNN model.Thereby the network model can be applied to more face scale situations and data set expansion is achieved.The experimental results show that the proposed multi-template multi-scale image pyramid data preprocessing method can improve the detection accuracy of dense multi-face.2)The idea of brain parallel interaction mechanism is introduced into MTCNN,and the subnetwork O-Net is modified for the task of detecting five landmarks on the face.Two sub-networks named O-Net-1 and O-Net-2 are proposed.The O-Net-1 makes the task of facial landmark localization work independently in the parallel interactive structure,thus improving the pertinence of the network to the task.The O-NET-2 introduces a parallel interaction structure for the facial landmark localization task and works independently with the original network's two-classes task and the regression frame detection task.Through experiments,it is shown that the MTCNN model based on the parallel interaction mechanism can improve the accuracy of facial landmark localization to a certain extent.3)A hybrid face detection model based on MTCNN model and ERT model is proposed.The model uses MTCNN as the main detection model,and ERT is used as the auxiliary detection model.The image not detected by MTCNN is sent to the ERT model for secondary detection.Through such a method to solve the problem of missed detection of a single network.Moreover,a more complex face data set was created according to actual needs.The effectiveness of the proposed hybrid face detection model is proved by comparing experiments on the data set.