Low Resolution Face Recognition Across Variations In Pose And Illumination

In recent years,with the improvement of computer performance and the widespread application of face big data and high-performance GPU computing.Machine learning and neural networks have achieved great success in the field of face recognition.However,it is not easy to truly integrate into our life,generally applied to real life security and surveillance videos.In addition to uncontrolled posture and illumination conditions,the low resolution of the face image captured by surveillance cameras is also a problem.Low resolution means less effective information is available and less easy to identify.The problem of face recognition for solving low resolution images is of great significance to the applicatio-n of face recognition algorithms in life.Combining with practical applications,this paper matches the facial image in the surveillance camera with the high-resolution frontal face image in the database.Convert the facial features of different resolution images into a new unified feature space,so that the distance between them is similar to the distance of the picture in the same resolution,pose,and illumination conditions,so as to achieve a low resolution face Recognition Across variations in pose and illumination.The paper cites the above theoretical study for program design and numerical experiments.The existing multi-dimensional scaling(MDS)is used to learn the mapping matrix from different resolution images to the unified feature space.In the experiment,the feature itself is used instead of the radial basis function to solve the problem.It reduces the computational complexity and highlights the performance of the unified feature space.After comprehensive comparison,we use MTCNN(Multi-task Cascaded Convolutional Networks)for face detection and alignment,and then using TCDCN for facial landmarks location,can get more precise results,eliminating the need for manual calibration.Using stereo matching tocalculate the similarity of the two images in the space takes too much time,so this paper uses simple cosine similarity,which reduces the amount of computation.The first part of the thesis mainly introduces the background and status quo of the face recognition algorithm,and the significance and content of the research topic.The second part introduces the overview of low-resolution face recognition and the comparative analysis of the main algorithms.In the third part,the theory research of multidimensional scaling(MDS)is introduced,the formula is deduced,and the dictionary learning model framework is given.In addition,the structure of the unified feature space is described in detail.The complete process of face public transformation matrix is learned by multi-dimensional scaling,formulas are derived and the algorithm steps are summarized.In the fourth part of the paper,the algorithm is implemented in the program.Experiments are performed on the Surveillance Camera database and the Choke Point data set respectively to show the experimental results.