Research On Source Camera Identification Based On ResNet And Multi-scale Feature Fusion Strategy

With the rapid development of today's society,the concept of artificial intelligence has been widely known in life.Statistical artificial intelligence which based on deep learning has also begun to shine in industry and academia.At the same time,the country has also begun to vigorously promote artificial intelligence in human life and society.As an important social medium,the importance of images can be imagined,and now the security of images becomes more and more important.The malicious tampering of the image content will not only cause public panic,but also reduce the credibility of the society,however the perpetrator will benefit from it.Therefore,the source camera recognition of the image is to some extent,a technique that protect the public from the harm of the devil.The traditional PRNU-based source camera model detection algorithm first extracts the PRNU pattern,and uses the correlation for pattern comparison to identify the given image when testing.This method has a lot of limitations and is susceptible to interference from the content of the picture.All pictures tested must be of the same size and cannot be cropped.In reality,the images that taken by different cameras often have different resolutions,so this is too restrictive for the application of PRNU-based method.The camera model recognition algorithm based on deep learning often extracts patches and then train the patches to identify them.This method only needs the same size of the patches,and does not require the resolution of the images themselves,so it is widely used.An algorithm based on ResNet and multi-scale feature fusion has been proposed in this paper,mainly uses the method of setting contrast experiment to verify the validity of the network for data collection and processing.First of all,for data acquisition,we propose a criterion function.Based on the criterion function,two methods of patching are adopted.The first method is to directly use the criterion function to score the all patches,and then sort the list of socres.only the first 64 patches will be extracted,the second method is sampling uniformly from the patches list which sorted by scores.Also,there is a third method,which does not use the criterion function,just directly performs random sampling.Secondly,for the data processing method,Gaussian high-pass filtering.LBP operator and no processing are used for comparisonWe used the above nine methods(3*3)to build nine datasets,and conducted a control experiment to train and test our model.In the case of using the same extraction method,the datasets processed using the LBP operator have the fastest rise in the accuracy of the first two epoch trainings,followed by the data processed by Gaussian high-pass filter,and finally the unprocessed data.But afterwards,the network trained by unprocessed data has the fastest convergence speed,followed by LBP,and finally the data processed by Gauss high-pass filter.It can be seen that the preprocessing has certain effects in the initial stage of training,but when a fixed filter kernel is used,the results is not good,because the low-frequency features that can be used for classification are suppressed,meanwhile LBP is a rotation-invariant feature.which has certain advantages for this problem.However,accounting for the strong feature extraction ability of deep learning,the use of LBP is not particularly large.For the same pre-processing method,different sampling methods have different emphasis.During the research,it was found that the distinction between the three Sony camera models are difficult,when processed by Gaussian high-pass filterorLBP,the three will be easy to identify,but this operation will decrease the performance of classification of other categories.