Classification Between Ps And Stego Images

For the present steganalysis which can not accurately distinguish the classes of images after treatment of image processing methods (referred to as the PS image), always recognize the PS images as natural or stego images by error, making the correct classification for images more difficult, especially might bring significant detection error for decision of the stego images. This thesis mainly goes along with researching for the problem of correctly classification of the natural, stego and PS images. The bodies and novelties of researches are outlined as follows:(1) Based on the correlations between the image content and the steganagraphy and PS processing technologies, corresponding additive noise models for images are constructed, which classify the stego and PS images into different additive noise models. On this basis, a novel classification frame for images is constructed.(2) For the noise model, an image classification algorithm is presented, which based on the characteristics of wavelet high-frequency decomposition. According to the differences among different classes show their differences of additive noise, and the property of the noise behaves in high-frequency, and the property of high-frequency wavelet decomposition is noise-sensitive. This thesis classified the natural, stego and PS images based on the features of high-frequency wavelet decomposition. The experimental results demonstrate that the high-frequency is sensitive to the noise, the detection results is well for the image classification.(3) Analyzing the images classification availability of the features which are based on wavelet de-noising that can remove the influence of the image content to classification, the features which combine the holistic features and the removed content features, the features which have lacal characteristic after images divided into blocks; And finally combining the three features of holistic features, local features and features that remove image content, then the images are classified. The experimental results demonstrate that association features based on different purposes and property reflections, the performance is higher.Finally, the thesis is concluded in the last chapter where the limitations of the proposed methods are discussed as well as possible countermeasures.