Research On Evaluation Method Of Image Visual Privacy-preserving Level In Compressed Sensing Coding Network

Nowadays,computers can analyze and process images faster and deeper,which means digital images can provide more information.People can use machines to automatically understand and analyze the content of images or videos to improve their daily life.But at the same time,people also hope that their privacy does not be violated.Therefore,how to strike a balance between the two is an urgent problem for many practical applications.For this problem,the reasonable analysis and processing for the visual privacy in the image can eliminate the image visual privacy threat to a certain extent,and image visibility and the understandability for machines can be maintained to meet the requirements.Therefore,to resolve the conflict between computer vision processing and image visual privacy-preserving,it is necessary to protect image visual privacy and evaluate image visual privacypreserving level objectively and accurately.The research on the evaluation method of image visual privacy-preserving level concerned in this thesis is to deeply study the relationship between human vision and image visual privacy,and finally let computers simulate the human visual system and human perception process to the image privacy.In this thesis,image visual privacy-preserving is implemented in a network framework based on compressed sensing and a computable image visual privacy-preserving level evaluation is performed on the images in the network,which can provide a guide for image recognition,classification and other tasks,and strike a balance between computer vision processing and image visual privacypreserving.This thesis first proposes a new idea for the image visual privacy-preserving,namely the compressed sensing coding network(CSCN).Each layer of the network obtains a lower-resolution image by dimensionality reduction projecting of the previous layer image to the image block based measurement matrix.Therefore,the CSCN with different depths can obtain images with different visual privacy-preserving levels.These images are still robust for computer vision processing tasks such as recognition and classification,achieving a balance between image visual privacy-preserving and image expression.In this network,a modified non-negative Gaussian(MNG)random measurement matrix is applied,which solves the problems that image feature loss and poor recognition performance caused by the general measurement matrices in the CSCN.In addition,a Gaussian fitting normalization method is also used to solve the problem of "overflow" of image pixel values in the network,which can restore images in the CSCN with high quality.This thesis also researches an image visual privacy-preserving level evaluation for the CSCN(VPLE-CSCN)method based on the colorfulness,contrast and salient structural features to analyze and evaluate the image visual privacy-preserving level of different layers in the CSCN.The basic concept is to use speckle noise adaptive weighting based colorfulness measurement(SNAW-CM),color/cube asymmetric alpha-trimmed mean enhancement(CAAME)and salient generalized centersymmetric local binary pattern(SGCS-LBP)operator to extract colorfulness,contrast,and salient structural features in the image,respectively.The features are fed into a support vector and fuzzy cmeans combinatorial regression model to obtain the final evaluated image visual privacy-preserving score.The appropriate network depth can be set according to the evaluation result and actual needs,which can retain useful image visual information to the maximum extent and achieve the goal of eliminating image visual privacy threat to varying degrees.Finally,this thesis verifies the recognition robustness of the CSCN and the coding effectiveness of the MNG random measurement matrix by recognition experiments on the ORL face database.The normalization effectiveness of the Gaussian fitting normalization method for grayscale and color images in the CSCN is also explored in this thesis.Based on the CSCN,this thesis constructs image quality and image visual privacy-preserving level evaluation databases LIVE?CS,TID2013?CS and CSIQ?CS with two labels of image subjective quality scores and image subjective visual privacypreserving scores.On these three databases,this thesis studies and analyzes the measurement effectiveness of the SNAW-CS and the CAAME,and the robustness of the parameters in the VPLECSCN,the effectiveness of image quality assessment,the reasonableness of prediction,the performance comparison with other methods,the effectiveness of extracted features,the computational complexity and the computational resource consumption of the VPLE-CSCN.Experimental results show that the CSCN has better recognition robustness under visual privacypreserving and the VPLE-CSCN has better prediction effectiveness.