Research On The Algorithm Of Image Privacy-aware Based On Deep Learning

The popularity of the Internet has brought tremendous changes to people's lives.Nowadays,more and more people prefer to sharing their information on social platforms.However,users are unconscious of the behaviors of sharing information sometimes will cause personal privacy leakage.In the Internet era,images are a common type of data,and the privacy in images are concealment and subjectivity.It is often difficult to detect whether privacy is contained in the image.Therefore,it is especially important to identify the risk of privacy leakage of image content.Image privacy-aware technology refers to the technique of extracting privacy features through image privacy-aware algorithm,and then using privacy features to construct a classifier to identify image privacy.Compared with the general image classification task,the image privacy-aware task has the characteristics of large intra-class gap and small inter-class gap.Therefore,it is more difficult to accurately characterize the privacy features of images by using traditional image features.In recent years,with the rapid development of deep learning technology,deep learning technology has shown great advantages in image privacy-aware tasks.Image privacy-aware methods based on deep learning have received more and more attention.This dissertation proposes a new method based on the character of image privacy-aware problems and the methods with deep learning in the latest development of image learning tasks.This method combines the deep learning method with the multi-label privacy attribute of images to improve the performance on image privacy-aware.In this dissertation,we improve the bilinear CNN network structure to make the model focus on image privacy features,and add multi-label privacy attributes to make the network learn objectives from a single privacy target to a specific image privacy category.Saliency maps are generated based on convolution features for locating privacy region,and the privacy region is selected by the HSV color space model.These above steps complete the positioning of the privacy region.During the study,we make a large number of comparative experiments to verify the validity of the image privacy-aware model.In this dissertation,in order to verify the robustness of the features extracted by the image privacy-aware model,we use the Sniffdata dataset with more than 9,000 images to analyze the performance of the model.The accuracy rate achieves 96.9% on this dataset.Moreover,we use other two public datasets,PicAlert and VISPR,to evaluate the generalization ability of the model.The accuracy rate achieves 87.7% on the PicAlert dataset,while the mean average precision(mAP)is 47.54%.Through analyzing the experimental results,the proposed improvement spots of image privacy-aware can effectively improve the effect of image privacy-aware model.