Research Of Large Pose Face Augmentation Andrecognition Based On 3D Face Reconstruction

The advancement of face recognition technology has been greatly promoted by the development of deep learning technology,and a lot of face recognition algorithms and deep convolutional neural networks have emerged in an endless stream,which get the state-of-art results in diffrenent datasets.However,all of these methods are researched on the specific datasets,and they are relatively simple,thus the face recognition accuracy of net trained on thest data will be vary low due to the influence of illumination?occlusion and background in the actual environmenta,which can not be applied.To solve these problems,this thesis studies face augmentation algorithms and face recognition algorithms of large-pose face,the main contents include facial point detection and face augmentation for large-pose face,and face recognition algorithm for these data.These algorithms can achieve effective face augmentation on single face photo,and these datasets are very similar to the one that in the real environment,which can be trained use the neural network to achieve effective recognition of large-pose face samples.The main work include:(1)Analyzed the importance of facial feature point detection in face recognition,and researching on large-pose facial point detection algorithm,which include algorithm based on cascaded convolution neural network and facial point detection algorithm based 3D face reconstruction,and the latter can effectively detect the facial point in various face pose,especially in the larger pose,the algorithm can obtaion smaller detection error.(2)Compared to the characteristic of large-pose faces,three face augmentation algorithms are researched in detail,which include the algorithm based on traditional image processing ? algorithm based on convolutional neural network and 3D face reconstruction?and the one that based on pixel transformation,and the contribution of the three algorithms to the face recognition accuracy has been analyzed.And the latter two algorithms are used to agument the multi-pose faces in a single real environment face photo.Experimental results show that VGG-16 network trained on these data can greatly improve the face recognition performance.(3)Analyzed the advantages of face feature extraction of three neural networks VGG-16,FaceNet and Center Loss in face recognition.Further analyzed the characteristic of each network in video surveillance face.The results show that the face augmentation algorithm based on pixel point transformation combined with FaceNet neural network can obtain the effective face recognition of large pose faces.