Research On The Privacy Protection Algorithm Of License Plate And Face For Video And Image

Visual privacy protection refers to the desensitization of sensitive information that appears in images and videos,while retaining the utility of the data,so that the data can be released,shared or used without exposing personal privacy.License plates and faces are common objects in videos and images,closely related to identity information,and are key privacy protection objects.Most of the existing privacy protections are based on data collected on the Internet,which contain few real visual scenes and insufficient labeling.In the video of real visual scenes,the environment is complicated,and it is a difficult task to accurately locate small targets such as license plates.However,the existing license plate detection algorithm has a low recall rate in real scenes,and the risk of license plate privacy leakage is high.In addition,the increasing advancement of face recognition technology also makes face images pose a great threat to personal privacy.Even desensitization methods such as mosaic and blur cannot guarantee the concealment of the identity,and there is a risk of being attacked by malicious users to restore the identity.This thesis has conducted an in-depth study of the above issues,and the main work is as follows:1)Aiming at the problem that the existing privacy protection dataset contains few real visual scenes and insufficient labeling,this thesis constructs the license plate privacy protection dataset(LPPD)and the face privacy protection dataset(Celeb A-P).LPPD is constructed based on surveillance videos collected from real scenes in a city in China.It contains several common real-world scenes such as traffic,community,parking,etc.It is annotated with license plate privacy attribute labels,as well as a variety of utility attribute labels such as vehicles,and the license plate privacy is classified into three privacy levels.Celeb A-P is constructed based on the Celeb Faces Attributes Dataset.In order to comprehensively evaluate privacy protection,this thesis also defines corresponding evaluation indicators from two perspectives: the effectiveness of privacy protection and the utility of desensitized data.2)Aiming at the problem of difficulty in privacy positioning of small targets in real scene videos and the low recall rate of existing license plate detection models,this thesis proposes a privacy location algorithm for license plate based on Kalman filter,namely VTKF.It introduces vehicle tracking to obtain multi-source data,associates the license plates between frames through vehicle id inheritance,and then integrates multisource data to build a novel two-way Kalman filter prediction model.This model uses the temporal and spatial cues in the video sequence to predict missed license plates and improve the recall rate of license plate privacy.In the process of prediction,VTKF uses the relative position of the license plate and the vehicle to constrain the predicted value,making the predicted value more accurate.Experimental results show that compared with existing license plate detectors on the LPPD dataset,VTKF successfully increased the recall rate by 8.5%,reducing the risk of license plate privacy leakage.3)Aiming at the problem of personal identification information leakage caused by the attack on the face image,this thesis proposes a face de-identification algorithm based on the generative adversarial network,namely Privacy GAN.It constructs a perceptual encoder by introducing identity authentication loss,generates deidentification latent code,and adds latent code for target attributes to explicitly modifies facial privacy attributes,while retaining the invariance of facial expressions,postures and other information that is irrelevant to privacy.Finally,a new face image that protects identity information is generated.Experimental results show that the algorithm effectively hides the identity while generating high-quality desensitization data.Compared with the existing methods on the Celeb A-P dataset,it increases the privacy score by 1.21%,and increases the utility score of the desensitized data by 4.86%,which optimizes the trade-off between privacy and utility.