Research On Key Issues Of Face Detection And Recognition In Random Video

Face recognition in video captured randomly always suffers from video jitter, variable lighting, pose, expression and so on, and illumination is deemed as major factor for recognition, as well as image processing and artificial check subject to video jitter. In this paper, we do research on video stabilization, face detection, face feature extraction and classification in video captured under variable lighting. The main contents are as follows:In chapter 1, the related study background, significance and purpose of face recognition are expounded, as well as difficulties and challenges in face recognition. On the basis of the current research on video stabilization and illumination for face recognition in video, the main research content and idea of this study are proposed.In chapter 2, generalized process of face recognition is introduced. Key technology of face recognition in video captured randomly is dissected, which is the basis of further study.In chapter 3, The problem and challenge of video stabilization is studied and analyzed in detail, and a new method based on modeling of motion imaging (MI) and adaptive step length motion filter is proposed to eliminate the fast scanning motion effect and extract jitter motion. The MI fuses the scanning motion of camera into the rotation+translation model to build a new model which is employed to eliminate the fast scanning motion. To meet the requirement for local motion, second-order moment of block is employed to estimate local motion which realizes to locate feature. Adaptive step length motion filter can check if the process of jitter finishes and separate jitter from the global motion. The experiment results show that the proposed algorithm is effective and robust in video stabilization under fast scanning motion.In chapter 4, face detection based on Adaboost is studied. Adaboost and Haar-like feature are combined to detect face in image or video in this paper. Space constraint between eyes region and face region is employed to eliminate the spurious face region.In chapter 5, High-order illumination invariant extraction is proposed based on filter and LBP. For the problem of estimated face illumination invariant is affected by illumination and filter parameters, we proposed to code the feature by LBP to extract high-order feature. The proposed method can not only weaken effect of illumination and parameters deep, but also weaken noise effect of LBP, and so the high-order illumination invariant is robust to noise. Experiment results demonstrate that the high-order feature is more robust to illumination.In chapter 6, decreased dimension illumination invariant extraction based on non-local NeighShrink denoise model is proposed. The proposed model takes the non-local means filter into NeighShrink denoise model to estimate face reflectance which is more robust to illumination. Meanwhile, data fusion is employed in first-level wavelet domain to obtain decreased dimension feature. The experiment results show that the proposed method can estimate reflectance more robust to illumination, and less computation in latter process.In chapter 7, ensemble based on improved plurality voting is proposed, and ensemble diversity is studied too. In improved plurality voting, confusion matrix is employed to estimate the probability at which the component classifier classifies the input as some class to improve the classification performance on multi-class classification. A new diversity measure based on Shapley value is proposed for the shortage of existing diversity measures. The proposed diversity measure realizes to build relationship between diversity measure and classification results. Experiment results demonstrate the improved plurality voting outperforms plurality voting on multi-class classification.In chapter 8, a face recognition platform is built, and it realizes auto face detection, feature extraction and recognition, as well as stabilization for jittering video.In chapter 9, the major of the study, conclusion and innovation are summarized. At the same time, further work is predicted for providing references for further research on this work.