Research On Video Face Recognition Algorithm In Unconstrained Environment

Video face recognition is a popular research direction in the field of machine vision.It is convenient and efficient when applied to occasions such as identity verification,and is widely used in the fields of justice,finance,security and so on.However,most of these applications are based on face recognition in a restricted environment with an ideal acquisition environment or staffing.Since face recognition in real life is easily affected by environmental factors such as lighting conditions,posture changes,and expression changes,video face recognition research in unconstrained environments is relatively challenging.The use of deep learning,especially the convolutional neural network algorithm for face recognition research,has achieved remarkable results in recent years.In this paper,the latest academic trends in the field of face recognition have been studied.It is found that deep learning based on convolutional neural network is the mainstream method for video face recognition research.The main reason for the impact accuracy is the noise caused by the complex and varied environment,such as changes in lighting,posture,and expression.For the face recognition problems of complex and variable environment such as changing posture and variable illumination,this paper uses VGG(Visual Geometry Group)deep convolutional neural network as the basic network,respectively introduces symmetric positive definite matrix and singular value decomposition to carry out the network.Improved to integrate facial information across frames to achieve efficient and robust video facial recognition.A symmetric positive definite matrix can model an image through local descriptors of a set of image sets.The singular value decomposition can reduce noise and compress the image matrix for images of different scales.The methods proposed in this paper are carried out on the internationally exposed face dataset(YouTube Face Dataset,YouTube Celebrities Dataset).The main work and achievements of this paper include the following aspects:(1)This paper combines a convolutional neural network with a symmetric positive definite matrix representation to design a new network structure to represent and classify image sets.The video is treated as an image set,and a matrix is used to represent an image set.The symmetric positive definite matrix models the image through local descriptors of a set of image sets to represent the relationship between the image and the image in the video.In addition,the use of normalization in the network structure furtherenhances the classification capabilities of the network.(2)Proposed convolutional neural network combined with singular value decomposition to represent and classify image sets.Singular value decomposition can reduce noise and compress matrix data.In this paper,the relationship between singular value and singular value after singular value decomposition is analyzed,and the singular vector corresponding to the larger singular value is used to reconstruct the image.The reconstruction results show that some singular vectors and singular values with large singular values can effectively express the characteristics of the image.In this paper,a deep convolutional neural network and singular value decomposition are combined to design a new network structure to classify and identify face video,and select different numbers of singular values and singular vectors to analyze different singular values and singular vector pairs.The impact of the experimental results.(3)In order to verify the effectiveness of the proposed algorithm,this paper conducts experiments on two internationally published databases YTC and YTF.By using the face recognition results based on pictures as the experimental reference,the method of this paper has significant advantages.The experimental results show that the recognition results of the video-based face recognition method compared with the face recognition results of the relative voting method of the single image increase the recognition accuracy of the two data sets by 1.06% and 3.31%,respectively.