Improved Planar Rotation Face Detection And Recognition Of Cascaded Convolutional Neural Networks

This paper firstly proposes a face detection algorithm based on improved concatenated convolutional neural network for the difficulty of face detection under various Rotation in Plane(RIP)angles.Experiments show that the algorithm can accurately detect faces under arbitrary plane rotation angles.Then combined with Facenet network and XGBoost classifier,the face image detected by RIP angle is recognized.Experiments show that the proposed method has higher accuracy for face recognition under RIP angle.The main work of the thesis includes:1.This paper improves the MTCNN cascade convolutional neural network.The MTCNN network has a total of three levels of 12 net,24net,and 48 net.The cascading convolutional neural network before improvement can only detect faces with small rotation angles in normal scenes.A RIP angle judgment is performed by embedding a classification network(Cnet)between the first stage(12net)and the third level(48net)of the cascaded neural network,and a parallel structure is formed by the second level network(24net)before the improvement.Cnet is a small convolutional neural network consisting of three convolutional layers and two fully connected layers.After each convolutional layer,the BN layer and the pooling layer are added.Cnet outputs four categories through the cross entropy loss function,and the plane rotated face images are divided into one class every 90° in the vertical direction.During the test,the face candidate box of 12 net output is classified by Cnet,and then subjected to affine transformation and input to 48 net for screening detection.2.In this paper,a face recognition method is proposed based on Facenet network and XGBoost classifier.The rotating face of the detector output retains the classification flag,and the face information is rotated into a positive face,and then Facenet is used for feature extraction to make the face image change.It maps a 1*128-dimensional embedded layer feature vector through the triplet loss function,and then inputs the feature vector into XGBoost training.When training XGBoost,you need to set a variety of parameters,the general parameters select gbtree,which represents the tree-based model for lifting calculation,and the task parameters use softmax to handle multi-classification problems.The lifting tree parameters were determined by observing the training loss curve.3.Extend the LFPW face data set to cover the Cnet network after covering all the rotation angles in the plane,and embed it into the MTCNN to form a plane rotation face detector.Then compare the algorithm and improve the detection accuracy of the preMTCNN on the FDDB dataset and the planar rotating face dataset made in this paper.The experimental results show that the improved method has a recall rate of 0.847 and an accuracy of 0.9 when detecting 853 RIP angle faces.In face recognition,download the PubFig face data set,select 2567 pictures of 29 people,divide training sets and test sets by 3:1,and compare the three classification algorithms after feature extraction.The experiment shows that the face recognition accuracy rate is above 92%.It also proves the effectiveness of extracting facial features.