Research On Median Filtering Forensics Based On Cosine Similarity And Local Binary Pattern

In the period of information digitization,the rapid development of image processing technology,the excellence of image editing software,and the tampering of images in the network through image processing software can be seen everywhere.When tampering images are used in military,political and other sensitive fields to achieve illegal purposes,the harm can involve the entire country or even the entire world.It can be seen that the authenticity of the verification image is of great significance.However,in image editing processing,as a smoothing filter,median filtering has good edge retention and nonlinearity.Image tampering masks the image tampering marks by median filtering,which makes median filtering widely used for information hiding and anti-forensics technology.In order to reveal the processing history of digital images,passive forensics of median filtering has become a hot research topic.The background and significance of image passive forensics is analyzed in this thesis,and combines the research status of median filtering forensics based on support vector machine,the traditional forensic algorithm based on support vector machine and the median filtering forensics algorithm based on deep learning convolutional neural network is also analyzed.The development status of convolutional neural networks is briefly introduced.In this thesis,three traditional median filtering forensics algorithms based on local binary mode are compared and analyzed.The existing deficiencies and research directions based on local binary mode median filtering forensics are obtained by simulation experiments on different data sets.In order to reveal the mathematical statistical relationship between image pixels before and after median filtering,the fusion feature CLF(CS & LBP Fusion Feature)for image median filtering detection is designed.As a linear vector correlation index,cosine similarity is combined with the local binary pattern feature to obtain the fusion feature CLF to perform median filtering on the image.The experimental results show that the fusion feature can effectively detect the median filtering of the original image and JPEG compressed image.In the calculation of the cosine similarity feature,the strategy of obtaining the block pixel matrix by using the local square block scan of the difference image reduces the calculation time of the whole feature and effectively decreases the time complexity of the feature calculation of the big data set.Combining with the advantages of deep learning convolutional neural network toautomatically learn image features,a local binary mode median filtering forensics algorithm based on convolutional neural network is designed.Different from the general image classification,the operation trace of the median filter is very weak in the original image and the filtered image,and can not be directly discerned by the naked eye.In order to make the network better learn the traces left by the median filtering in the image,the algorithm adds the local binary mode preprocessing operation before the data input convolutional neural network.An LCNN network with an LBP pre-processing layer is designed.Firstly,by calculating the LBP binary pattern feature map of the original image and the corresponding median filtered image as the preprocessing of the LCNN,the two types of image differences are amplified.Then,the LBP binary pattern feature map is put into the general convolutional neural network for feature learning and classification.The experimental results show that for the image datasets of JPEG compression,the inter-class difference is still valid after LBP pre-processing.The LCNN network can achieve effective detection on both data sets and achieve higher detection accuracy.