Font Size: a A A

Research On Visual Privacy Protection Technology And System Of Service Robot

Posted on:2022-01-13Degree:MasterType:Thesis
Country:ChinaCandidate:J C LinFull Text:PDF
GTID:2518306527469434Subject:Mechanical engineering
Abstract/Summary:PDF Full Text Request
Visual privacy protection is an important issue in the process of social development.To address the problem of human privacy leakage caused by social robot vision equipment from the source,we study from three aspects that full-body visual privacy recognition,fullbody visual privacy protection and face visual privacy de-identification based on generative adversarial networks(GAN)for social robot,respectively.In order to solve the problem that social robots can not effectively recognize the visual privacy behavior,a visual privacy behavior recognition algorithm based on GAN is proposed.To overcome the problem of user privacy leakage caused by social robot's vision devices,we proposed an image privacy protection method of social robot vision base on improved CycleGAN(Unpaired Image-toImage Translation Using Cycle-Consistent Adversarial Networks).Aiming at the problem that facial visual privacy protection can not be effectively protected,this paper proposes a facial visual privacy protection algorithm based on GAN.Based on the above situation,this paper studies the visual privacy protection system of home social robot,the main contribution and innovations of this paper are as follows:1)A visual privacy behavior recognition algorithm based on GAN is proposed.In the traditional visual privacy recognition process,it needs many labeled data,which increases the cost of deploying the privacy recognition model.To solve this problem,this paper proposes a visual privacy behavior recognition algorithm for social robots based on improved GAN.Among them,the 9-layer residual network improves the quality of the image generated by the model generator,while the discriminator structure with 10-layer network enhances the feature extraction.The objective function,the loss function,and the strategy of dynamically adjusting the learning rate designed for the visual privacy behavior recognition algorithm of social robot can ensure high performance.2)We proposed an image privacy protection method of social robot vision base on improved CycleGAN.The traditional visual privacy recognition model can't deal with the leaked visual privacy image effectively,while the traditional visual privacy protection model can't get the high-quality visual privacy de-identification image.To solve this problem,we propose a visual privacy protection method for social robots based on improved CycleGAN.Firstly,focus on the generalization ability,an improved CycleGAN algorithm was proposed by optimizing the structure and parameters of CycleGAN.Then,a training dataset growth algorithm based on normalized cross correlation was proposed,which is used to automatically enrich the training dataset according to the environment change.Finally,the model was implemented on our developed social robot,and then the source domain and target domain training dataset with 48,000 images were established,and the variate datasets based on 6 kinds of test solution were constructed.3)A facial visual privacy de-identification algorithm based on GAN for social robot.The traditional face image generation algorithm can not protect the visual privacy effectively.Aiming at the problem of face visual privacy protection,this paper proposes a face visual privacy protection algorithm for social robot based on GAN.Firstly,its convergence is proved mathematically.Then,an improved U-Net generator is used to improve the quality of the generated image,and two 7-layer network discriminators are designed to enhance the feature extraction ability of the model.Then,we propose pixel loss,content loss,adversarial loss function and optimization strategy to ensure the performance of the model.In this experiment,we apply the algorithm to the face visual privacy protection of service robot,and analyze the relevant conditions that affect the effect of the model.Experiments show that this algorithm can effectively protect the human face visual privacy,and improve its robustness and effectiveness.4)A social robot visual privacy protection prototype system is proposed.Through the construction of the experimental platform and the organic integration of the above algorithm model,this paper proposes a service robot visual privacy protection model,obtains a set of service robot visual privacy recognition and protection system scheme,and then forms a set of home service robot privacy protection system with the functions of home environment detection and visual privacy protection.
Keywords/Search Tags:Visual privacy protection, Deep learning, Generative adversarial networks, Robot vision, Social robot
PDF Full Text Request
Related items