The Research For Digital Image Blur And Resampling Detection

With the universality of digital image acquisition devices and the popularity of powerful and simple image processing softwares, people can edit the image easily to beautify the image and enhance visual effects. On the one hand, it meets people's needs. But on the other hand, it makes it easier to tamper the image. In fact, the events of image forgery happened frequently in recent years. It shows the phenomenon of image forgery has been increasingly serious. Therefore, identification of the authenticity of digital images has been an urgent problem.Digital image forensics is a new technology which aims to identify the origin, authenticity and integrity from analyzing the image statistical properties. Image resampling is often used in synthesis tampering, and image blur is often used as the post-processing operation. In this paper, I focus on researching the technology of resampling and blur detection. After making the detailed theory analysis about present research achievement, I make a summary about present limitations. The most important is that I propose two new algorithms for blur detection. The main content of this paper includes:First, the paper gives a systematic introduction about research background, significance and current status of digital image forensics. Then, the necessity of my research is proved by analyzing the application of blur and resampling in image tampering.Second, the paper introduces image blur and resampling technology in detail and analyzes their effect on intrinsic characteristics of the image. At present, domestic and foreign scholars have proposed their algorithms in these two areas. I do some work about their typical algorithms, including theory analysis, programming and performance analysis. After making a summary of present problems, the paper looks into the future research direction.In addition, this paper proposes two new blur detection algorithms based on DCT domain. One is based on distribution feature of DCT coefficients, and the other one is based on joint density feature of DCT coefficients. The feature of first method is one-dimensional. We can get the result by setting a reasonable threshold and making the threshold comparison. The feature of second method is 225-dimensional.We use SVM for training and classification. The two algorithms can resist the JPEG compression. In particular, the second algorithm shows its good adaptability to BMP, TIF and JPEG format images.