Research On The Detection Methods Of Natural Images And Computer-generated Images

The popularity of digital images has made the image data more prominent as away of information transmission. With the development of computer technology, thecomputer software is able to create vivid and lively computer graphics, which arevery similar to natural images. As a result, it is difficult to tell the difference betweenthem by our naked eyes. These images largely enrich people’s daily life, meantimemeet the needs of commercial advertising and magazine cover. However, it alsobrings the following questions. If the computer-generated images are mistaken fornatural images and applied to the court evidence and political propaganda, it willweaken the credibility of society and cause adverse social effect. Therefore, it isgreat significance to authenticate whether the scene of the digital image is authentic.This paper extracts higher-order statistical features from different angles, on thebasis of research of the statistical properties of images, and differentiates thecomputer-generated images from natural images without relying on anypre-embedded or pre-signed information. The technical research innovationsobtained are described as follows.First, the proper selections of the wavelet transform objects, wavelet basis andthe order of wavelet transform impact on test result are studied. To analyze the effectiveness of wavelet high-order features, we firstly study the impact of thelow-frequency sub-band, high-frequency and sub-bands of different orientation onthe detection rate, then we research on the impact of the wavelet transform objectsand the color space components on the high-order features, and finally the relevantconclusions are drawn.Second, compared with wavelets, multiwavelets have compacter support andmore vanishing moments. What’s more, they can both meet the orthogonality andsymmetry. In practice, they can combine the important smoothness, compactsupport and symmetry perfectly. Therefore, a new detection algorithm based on themultiwavelet transforms is put forward.Third, based on the defects of computer graphics in the edge construction, aneffective method based on image edge features is proposed for identifying computergraphics, which combined wavelet higher-order features describing image edgedifferenceswith wavelet higher-order characteristics of the predicting error images.Fourth, based on the differences of pattern noise between natural images andcomputer images, this paper advanced a new detection method which combinedSNR features reflecting the noise difference more intuitively with wavelethigher-order characteristics of the predicting error images.