Research Of Face Recognition On Video

Research of video-based face recognition has broad application prospects. This paper focuses on three points:video initial screening, face representation and recognition algorithms.To complete initial screening of video information and remove redundant duplicate information, some methods of video segmentation were studied, including pixel-based method and color histogram-based method. Segmentation threshold was determined through experiments, comparing two methods’ precision and recall, finally effect of sample interval on video segmentation is studied.In order to get a better face representation, a precise human eye location method based on PCA and SVM was studied, presenting a single discrete recurrent neural networks for solving quadratic convex optimization problem (problem of SVC). Then the model’s ability to get the optimal solution and its convergence under random initialization was proved theoretically. After getting the Gabor feature, a selection model based on PCA and LDA dimension reduction method was presented.In order to get better face recognition algorithm, a variety of face recognition models were studied, including sparse representation model such as SRC, LASRC and collaborative representation model such as CRC-RLS, then a single collaborative representation classification model(SCRC) was presented. After that Joint Bayesian model was studied and an improved Joint Bayesian model due to the imbalance of training samples was proposed. Then correct rate’s relationship with the number of storage and the number of samples in each library were studied. Finally, a multi-to-multi face recognition model and its stability was studied, considering multiple comparison methods, experiment results implied that this model increases the correct rate significantly.Research was conducted mainly on two face database. One huge database was collected from the Internet, containing complex faces that are closer to real application scenarios. Another one is the PIE database, which is widely used in the field of face recognition. This paper focuses mainly on the 1:N mode, so in addition to the recognition rate of the first indicator, Rank 1～Rank 10 indicators are also adopted. Finally the algorithm’s performance under 1:1 mode are presented.