Research On Forensics Of Image Blur Operation

Digital image is widely used in news media,judicial authentication,court forensic and other fields.However,with the development of information technology,the popularity of smart phones and cameras allow more and more powerful digital image editor enable to be developed and used.More users can freely process and modify the image,with playing into some malicious user's hands.Hence,digital image is difficult to guarantee the authenticity and integrity.Research on image blurring operation forensics can reveal the history of operation and verify the authenticity and integrity of image data through technical means.Based on traditional feature method and deep learning algorithm,this paper researches on blur operation forensics.The main works include:(1)Due to the emergence and application of Single Lens Reflex(SLR)cameras make photos formed into blur inconsistency inevitably.For this kind of defocused image,the existing blur forensics algorithm is difficult to identify the splicing and locate the tampering area accurately.To solve this problem,this paper proposes an algorithm of blur forensics and splicing localization in defocused image.Firstly,we strictly define the raw naturally blur in the defocused image and the artificial blur in the splicing process.By analyzing the characteristics of blur operation,we propose a joint feature set based on posterior probability map,noise histogram and derivative co-occurrence matrix.Then the proposed feature set of multiple cues is used to design the locating scheme of defocused splicing image.The experimental results show that the proposed feature set of multiple cues is superior to the existing blur detection features in blurring detection and can accurately classify both parametric and non-parametric blur.Also we can expose the splicing region in the JPEG compression post-processing.(2)Detecting the order of multiple operation is challenging,because the latter operation is easy to interfere with the previous operation,which makes the forensics feature based on the known artifacts design of the previous operation invalid.To solve this problem,the existing multi-operation forensics algorithms are firstly grouped into three categories: detecting a specific operation in a fixed operation chain,detecting any chain in a chain with the specified operation,and detecting the order of multiple operation in a chain with the specified operation.Secondly,in the same operation chain,by analyzing the interaction between the blurring and resizing operations of different sequences,we design the forensic features of the order of multi-operation detection and propose feature extraction and detection algorithm.The experimental results show that the proposed feature has a great advantage in the detection accuracy under the consideration of the full parameter combination,and the detection results are better than the state-of-the-art methods.(3)When JPEG compression is used as the post-processing operation in the image tampering process,artifacts of the actual main tampering operation will be weakened or even eliminated,which makes the main operation to be detected difficult under the influence of post-processing.To solve this problem,we design a multi-operation robust forensics network based on deep learning method to research on a multi-operation chain composed of blurring and resizing.In the spatial domain feature extraction subnet,the preprocessing layer uses two high-pass filters to generate the corresponding neighborhood residuals,and then constructs the forensic features through the high-order feature extraction layer.In the frequency domain feature extraction subnet,Gabor transform layer and Gabor feature extraction layer are proposed to obtain the operation artifacts in frequency domain.Combined with the extracted spatial and frequency domain features,the design of MCRF-Net is completed through the following classification network.The experimental results show that the proposed multi-operation chain robust forensics network can accurately identify multi-operation chains,and is superior to the state-of-the-art operation forensics network in terms of model convergence speed and final classification accuracy.