Research On Face Detection And Tracking Algorithm Under Complex Background

As an indispensable part of face recognition system,face detection and face tracking have attracted great attention in the academic field.The research on face detection and tracking in a single environment has made a breakthrough.However,due to many uncertain factors in the complex environment,such as illumination change,occlusion,face angle change,image resolution,the performance in the complex environment is still poor,which will limit the performance in varying degrees.In this paper,the face detection and tracking algorithm under complex background is studied.The work is organized as follows:(1)Research on a coarse-to-fine facial detection algorithm.The existing general target detection models Faster R-CNN and YOLOv3 are studied,and they are improved to be suitable for face detection tasks.After comparing the advantages and disadvantages of the two models,a coarse-tofine face detection algorithm is proposed.The algorithm completes two classification tasks and two cascading position regression tasks of the face proposal regions by constructing a deep convolution neural network which includes three modules: feature extraction,face frame coarse location and face frame fine location.The experimental results on the WIDER FACE dataset show that the detection rate of this algorithm is 76% in complex background,which is equivalent to the accuracy of Faster R-CNN,but the detection speed is 10 times faster.(2)To develop a neonatal face dataset.The dataset needs to process the image or video of the neonatal,and then make the label.The dataset contains images with complex background,occlusion,multi-angle and multi-expression.The database is used to simulate the detection algorithm and get the detector suitable for the facial features of the neonatal.The experimental results show that the improved YOLOv3 detection is the best of the three models,and the K-means clustering is used to obtain a priori frame which is more consistent with the face size of the newborn,and the detection accuracy is 99.43%.(3)Research on face-landmark with multi-task.The five key points of face are predicted by using multi-task learning on the basis of the coarse-to-fine face detection algorithm.The five key points are the center of left eye,the center of right eye,the nose tip,the left corner and the right corner mouth.The face detection results and key points are used to correct the face.On the one hand,the corrected face can reduce the difficulty of the next operations in the face recognition system;on the other hand,the experimental results show that adding key point detection can improve the face detection rate by 0.5%,indicating that the key point detection plays a certain role in supervision and improves the face detection quality.(4)Research on the tracking algorithm combining Siamese Network and Region proposal Network,and applies it to neonatal face tracking.The initialization information of the tracker comes from the result of face detection of the first frame image of neonatal video sequence by YOLOv3.It is found that adding more difficult samples with more semantic information can improve the discrimination ability and robustness of the tracker by comparing the tracking effect of the tracker obtained from different training sets.