Research On Key Technologies Of Visual Privacy Protection

With the rapid development of computer science and technology,the access of information has become increasingly diversified,and human's life has become more autonomous and independent.The popularity of video personalization services and the widespread use of the Internet have brought great convenience to people's daily lives.However,the real-time access to information has made people's privacy threatened.The personal privacy information in videos and images is the mainstay.The study of visual privacy protection methods has played a decisive role in solving personal privacy and security issues.Face de-identification is one of the most common visual privacy protection methods.It can provide effective face information while protecting personal privacy information,which greatly meets the requirement of practical application.However,in the traditional face de-identification methods,when the privacy protection is performed,the effective information of the face,such as an expression,is changed,and the utility of face data in applications such as video live broadcasting is reduced.Therefore,research on expression preserved in visual privacy protection is of great significance and application value for balancing privacy protection and data utility.In order to solve the problem that face expression may be changed when protecting visual privacy,this thesis firstly proposes a visual privacy protection method with expression preserved based on Multimode Discrimination Analysis which uses the best Fisher standard to effectively generate independent orthogonal subspaces,each of which does not affect others.This thesis uses AAM modeling to normalize the face images to remove the influence of angle,pose,etc.The face images involving various attributes such as gender,race and expression are decomposed to obtain the orthogonal identity space by the Multimode Discriminant Analysis method,each of which presents a corresponding average feature face and parameter.These characteristic parameters can affect the related feature information.Changing the gender and race and other characteristic parameters while making no changes to the expression feature parameters can achieve the expression preserved privacy protection.The experiments show that the visual privacy protection method based on Multimode Discriminant Analysis has a good performance on the processing of face images with multiple features.At the same time,our algorithm can effectively protect privacy while keeping the expression unchanged.Then,based on the research of the traditional face de-identification methods,this thesis also proposes a visual privacy protection method with expression preserved based on K-Same algorithm.To solve the problem that the average value of k similar face images is used to realize the privacy protection while the facial expression may be changed in K-Same algorithm,this thesis chooses the k-nearest neighbor classification to select images with same expression feature extracted by complementary LBP/LPQ feature operators,and the k most dissimilarity images with the same expression are averaged as the de-identification image to protect privacy in this thesis.Experiments show that our algorithm has a better effect than K-Same algorithm in privacy protection,and expression can be effectively maintained.