Blind Separation For Permuted Image And Its Application

Blind sources separation (BSS), which became an active research topic in signal processing and neural network in the last decade, consists of separating a set of unknown signals from a set of linear mixtures of them, when no knowledge is available about the mixing coefficients. There are many potential exciting applications of blind sources separation in science and technology, especially in image processing and speech recognition. According to some hypotheses on the number of sources and the number of observed signals, BSS is divided into overdetermined blind sources separation (OBSS), welldetermined blind sources separation (WBSS), underdetermined blind sources separation (UBSS) and single channel blind signal separation (SCBSS). The SCBSS is comparing with multiple channel blind separation, which has more challenging. In the earlier research of the SCBSS, generally, most of the methods to SCBSS assume that the mixing mode of individual source signals is superposition. However, if the mixed mode is permutation, a quite different one from superposition, the approaches mentioned above will no longer be functional. In this dissertation, we investigate the BSS problem of the permuted image and introduce the thought into image tamper detection. The primary contributions of the dissertation are summarized blow:1. A first and in-depth study on the BSS problem of the permuted image. The mathematical model of the permuted image is obtained by its definition. According to the permutation mixing matrix is a special binary matrix, of which each entry equals 0 or 1, a BSS model of the permuted image is described by estimating the mixing matrix. For different source images, a series of single-channel blind separation algorithms are proposed based on parameters estimation.2. When the sources are JPEG images, a blind separation method based on compression factors estimation is proposed for permuted JPEG image. The permuted image is compressed again and the primary compression factors are estimated by calculating the correlation coefficients of image blocks with before and after recompression. Using the estimated compression factors, a mapping space is constructed. The permutation mixing matrixes can be accurately estimated by classifying the parameters in the mapping space, thus the source images can be separated. At last, the proposed method is used in JPEG image tamper detection. Simulation results show the validity and robustness of the proposed algorithm.3. When the sources are blurred images, two novel single-channel blind separation algorithms for permuted blurred image are proposed by using blind restoration. For permuted defocus blurred image, the defocus blur radius is estimated by the characteristics of permuted image in the frequency domain, and then the permuted image is restored by performing the Lucy-Richardson(L-R)blind restoration method. The ringing effect of restored image is measured by defining the sum of absolute pixel gradient, and the permutation mixing matrixes can be accurately estimated by classifying the ringing effect of each sub-block, thereby separating the source images. For permuted motion blurred image, both the motion direction and length of point spread function (PSF) are estimated by Radon transformation and extrema detection. Using the estimated blur parameters, the permuted image is restored by performing the L-R blind restoration method. The permutation mixing matrixes can be accurately estimated by classifying the ringing effect in the restored image, thereby the source images can be separated. At last, two proposed algorithms are applied to image passive forensics. Simulation results show the tamper regions can be accurately detected.4. When the sources contain interpolated images, a novel blind separation method based on interpolation factors estimation is proposed for permuted interpolated image. The periodic property of difference sequence is detected by finite-difference for permuted image, according to the periodic property, the various interpolation factor could be identified. Using the estimated interpolation factors, a mapping space is constructed. The permutation mixing matrixes can be estimated by classifying the parameters in the mapping space, thus the source images can be separated. At last, the proposed method is used in interpolation image tamper detection. Simulation results show the validity and robustness of the proposed algorithm. Compared with existing ones, the algorithm has simple principle and small computational load.5. For blurring permuted images, a single-channel blind separation scheme using double blur correlation is proposed. The permuted image is blurred again and the correlation coefficients of image blocks are estimated before and after double blurring. Using the estimated correlation coefficients, a mapping space is constructed. The permutation mixing matrixes can be accurately estimated by classifying the parameters in the mapping space, thus the source images can be separated. At last, the proposed algorithm is applied to image tamper detection with blurring. Simulation results show high detection accuracy for tamper images with various blurring operations. The proposed method has good robustness against Gaussian noise and lossy JPEG compression.