Research On Face Detection And Segmentation Based On Deep Learning

The acquisition of face information has always been an important research topic in the field of computer vision.With the rapid development of artificial intelligence technology,as an important identity information,face information has been widely used in fields such as identity authentication,human-computer interaction,and public safety.Face detection is the first and key step in face-related applications(such as face recognition,facial expression recognition,face super-resolution reconstruction,face pose correction),and its efect directly affects the application of subsequent related technologies.Therefore,the importance of face detection is self-evident.In the existing mainstream research work,face detection mainly realizes the classification of the face and the positioning of the face bounding box.However,in the face bounding box of the detection result,the face information may only occupies a part of it.The background image in the bounding box brings redundant information,so the extracted face features have problems such as background noise,rough spatial quantization and large dimension.These problems lead to some practical application of face-related technology is limited.Aiming at the above problems,this paper focuses on acquiring more accurate face information by combining face detection and segmentation in the same network architecture,and proposes a face detection and segmentation method based on deep learning.The main research contents of this paper is as follows.1)A new data set with face detection and segmentation annotation was constructed.5115 images were randomly selected from the dataset of FDDB and ChokePoint,and then annotated them with masks labels by the VGG Image Annotator(VIA)image annotation tool,which can be used for our model training.2)In view of the problem that the current mainstream face detection methods only realizes the localization of boundary box,which leads to the background noise and unsatisfactory detection accuracy of acquired face feature information,a face detection and segmentation method based on Mask R-CNN is proposed.In this method,region of interest are generated by combining resnet-101 with RPN network,and pixel level feature point positioning is realized by RoIAlign algorithm to improve the positioning accuracy of feature points.Finally,the binary mask of face is generated by full convolution network to realize the segmentation of face segmentation of face information and background in the image.This method introduces the segmentation operation into the traditional face detection tasks,integrates the face detection and segmentation tasks into the same network architecture,and realizes the end-to-end face detection and segmentation effect.3)To solve the problem that the face detection and segmentation method based on Mask R-CNN has low detection accuracy in multi-target face detection and small-scale face detection,a face detection and segmentation method with multi-scale feature fusion based on generalized intersection over union Mask R-CNN(MG-Mask)is proposed.In this method,the generalized intersection over union function is used to replace the traditional smooth L1 loss function to improve the detection accuracy of multi-target face.In addition,multi-scale feature fusion were added to FPN(Feature Pyramid Network)to improve the performance of small-scale face detection.The experimental results on the face datasets FDDB,AFW and WIDER FACE show that the MG-Mask can effectively improve the performance of multi-target face detection and small-scale face detection.