Research Of Face Detection And Facial Landmark Localization Based On Deep Learning

As a direct and reliable biological feature,human face has a wide range of applications in the fields of information security,social public security,virtual reality and augmented reality,and so on.Face detection and facial landmark localization are the key steps in face image analysis and processing.The accuracy of detection and localization will greatly affect the performance of various tasks such as face recognition,expression analysis,head pose estimation,gaze direction analysis,fatigue detection,three-dimensional reconstruction and animation synthesis.In recent years,deep learning methods have made breakthrough performance improvement in the field of computer vision.Therefore,in this paper,we focus on the research and algorithm design of face detection and facial landmark localization,which has important theoretical significance and practical value.As the first step of the face analysis and processing task,in order to improve the real-time performance and reduce model parameters of face detection methods,we designed a face detection module based on the single-stage object detector.To achieve high speed as well as maintain high performance,we extracted multi-scale facial features by using Inception module and multi-detection branches.Moreover,we used anchor densification strategy to increase the number of anchors for small faces and feature pyramid method to enhance the features extracted from the front detection branch,which effectively improved the detection performance of small faces.Finally,we made quantitative and qualitative performance evaluations on the benchmark face datasets and compared our face detection module with various face detection methods.In order to make full use of the inherent spatial relationship between facial landmarks and improve the localization precision in large head poses,we designed a facial landmark localization module based on the hourglass network.By using landmark heatmaps and extracting multi-scale features through repeated bottom-up,top-down processing,we effectively improved the facial landmark localization precision.To solve the problem of inconsistent definition of 2D and 3D facial landmarks,we trained our facial landmark localization module on both 2D and 3D facial landmarks labels,and achieved high localization precision on 300W-2D and AFLW2000-3D.We evaluated our face detection module and facial landmark localization module on the benchmark face datasets.Our face detection module had a light-weight(?5M parameters)yet powerful network structure.The face detection AP on the AFW,PASCAL Faces and FDDB dataset was 0.987,0.969 and 0.963 respectively,and the average runtime of the above test datasets on a single GTX 1080Ti GPU was 12ms/image.The normalized mean error(normalized by pupil distance)of our facial landmark localization module(2D)was 5.02 on the 300W fullset,4.67 on the 300W-Common subset,and 7.29 on the 300W-Challenging subset.The NME(normalized by bbox)of our facial landmark localization module(3D)for faces with different yaw angles was 2.56(0-30 degrees),3.31(30-60 degrees)and 4.42(60-90 degrees),and the mean NME was 3.46.Compared with other state-of-the-art algorithms,our face detection module achieved high detection accuracy while improving detection speed and reducing model parameters.In addition,our facial landmark localization module can extract face contour and 3D pose features at the same time,which effectively improved the localization precision for faces in large pose.