Research On Computer-generated Images Detection Technology Based On Fractal Dimension

With the development of information technology, the storage media for resources has been transformed from traditional paper to electromagnetic form. As a kind of information resources, image has been developed with the information technology. The traditional film image has been increasing replaced by digital image. Yet comparing to traditional image, digital image is more vulnerable to tamper and difficult to detect. Therefore digital image identification has become a difficult problem in information security field. Currently there are two main methods: positive image identification and passive image identification. In the passive method there are two aspects: identification of natural image, identification of computer-generated image. This paper is focus on the identification of computer-generated image.Computer-generated image is obtained by vividly reproducing the real-world scenes. The computer first constructs a geometric representation of the scenes, then simulates the real objects’physical attributes, such as object shape, optical properties, relative positions between objects, and so on. Similar to the identification of natural image tamper detection, the identification of computer-generated image includes two aspects: extracts image characteristics and image identification based on the characteristics.In recent years, public research has been focusing on the integer-dimension method. Main research achievements include: In space domain, abstract the mean, variance, deviation, kurtosis from image’s transform domain. And then classify these characteristics by machine learning. This kind of identification methods does not have a very effective detect rate.This paper investigates the computer-generated image identification in the fractal-dimension domain. Being different from conditional algorithm, the algorithm we propose takes advantage of the surface texture feature and self-similarity differences between these two kinds of images, extracts the local fractal dimensions for image blocks to distinguish the computer-generated images and photographic images.The objects in real world usually have a rough surface and complicated geometric structure, which are difficult to be represented by single model. On the other hand, computer-generated images are usually produced by cycling basic geometric elements such as point, line, plane, as well as non-geometric elements such as gray scale, color, line type, line width, therefore having very strong self-similarity. Due to the difference of object model, computer-generated image usually has more clear texture characteristics and stronger self-similarity. These differences can be revealed from fractal dimension. Based on the fractal differences between natural image and computer-generated image, the algorithm proposed in this paper first sieves the image blocks, removing the blocks with smooth variation, then calculate the fractal dimension of the remaining blocks. Then use these fractals as the image’s characteristic vector, and finally put the vector into SVM classifier for identification. Besides, while calculating fractal dimension, we use an improved box-counting method, which increases the accuracy as well as efficiency. The experiment indicates that this fractal dimension based algorithm shows a new direction for detecting photorealistic computer-generated images. It boasts a really satisfying detect rate and a promising study prospect as well.