【作者】 段新涛；

【导师】 方勇；

【作者基本信息】 上海大学 ， 通信与信息系统， 2011， 博士

【摘要】 盲分离是现代信号处理领域的一个新的研究热点,在诸多领域有着广泛的应用。单通道盲分离作为盲分离研究的一个前沿研究正日益受到关注。本文研究在图像篡改等实际应用中常遇到的一种特殊的单通道混合方式即置换混叠方式的盲分离问题。由于与传统的叠加混叠方式的单通道信号的混合方式不同,用以往的单信道盲分离的理论和方法很难解决,本文围绕置换混叠图像在特征域上的盲分离方法展开研究。置换混叠图像是由不同的源图像非交叠的置换而成的,使得置换混叠图像失去信息的完整性。源图像的来源不同,构成的置换混叠图也具有多样性。同时置换图像的位置、大小和个数都是未知的。检测和分离出置换图像的关键是找到置换图像和被置换图像之间特征差异,即找到显现特征差异的特征域。本文针对不同类型的置换混叠图像,在不同的特征域提出了置换混叠图像盲分离方法。论文的主要工作如下:1、在置换区域与被置换区域的图像存在形态差异的条件下,利用形态差异图像在不同字典上表示时有大的稀疏度差异,针对一类存在形态差异的置换混叠图像,提出基于稀疏分解的盲分离算法。首先,根据这类置换混叠图像的数学模型,得到其稀疏表示的模型。根据置换区域与被置换区域图像的特点,选择Contourlet基字典和Local DCT基字典作为特征域。分别迭代地在两个不同的特征域上进行稀疏分解,利用块定位松弛算法求得不同形态图像最稀疏的表示,从而实现置换图像分离。实验结果表明,对于一类包含分片光滑和纹理成分的置换混叠图像,本算法能够较好的把纹理图像从置换混叠图像中分离出来,同时,对纹理图像的大小、位置、个数和类型具有鲁棒性。2、置换混叠图像中的被置换图像和置换图像的来源常常是不同的,这些图像在获取、传输和处理等环节常常感染上噪声,从而使得置换后的图像的不同区域存在着噪声差异。针对这类在被置换区域和置换区域之间存在噪声差异的置换混叠图像,提出一种基于噪声检测置换混叠图像盲分离算法。由于学习字典具有高的自适应性,其稀疏表示的稀疏度比参数字典高,选择学习字典作为特征域。通过对置换混叠图像本身采样,利用非零元个数约束的K-SVD字典学习算法得到图像的稀疏表示,以达到去除噪声的目的。并与原图像求差运算来检测含噪置换图像所在的区域和大小,并通过图像形态学操作和设定阈值来实现其分离。实验结果表明该算法能只利用两个确定的值对噪声区域有效的去噪,并有效的分离出不同位置、大小、个数的置换图像。3、针对一类存在运动模糊差异的置换混叠图像,提出基于四向差分的运动模糊置换混叠图像盲分离算法。首先,在分析运动模糊的特征后,选择空域作为特征域。对置换混叠图像做四向差分运算,得到四个方向上的差分图像。计算每个差分图像子块的方差值并阈值化得到一个二值图像,以检测置换图像的大致区域。然后利用图像形态学操作优化二值图像,并乘以置换混叠图像以分离出置换图像。实验结果表明该算法有效的分离出不同位置、大小、个数的运动模糊置换图像,并不受运动模糊方向的影响。4、针对一类存在高斯模糊差异的置换混叠图像,提出基于差分进化的高斯模糊置换混叠图像盲分离。首先,在分析高斯模糊的特征后,选择空域作为特征域。把置换混叠图像的梯度图像分成小块,并为每个小块设定一个阈值,组成一个阈值向量。利用差分进化算法求得一个最优的阈值向量,利用优化的阈值向量把置换混叠图像阈值化为二值图像,并使用图像形态学操作优化二值图像,利用优化后的二值图像分离出置换图像。实验结果表明,与固定阈值法相比较,本算法能较好的把置换图像从置换混叠图像中分离出来,而不受置换图像位置、大小和个数的限制。更多还原

【Abstract】 Blind separation is new research direction in modern signal processing field, which has been applied to engineering. There is growing concern on single channel blind separation as frontier of blind separation research. However, in the applications such as image tamering etc, a special single channel alias mode called as permuted alias mode is encountered, which is different from the traditional superposition alias mode. Previous theory and method of single channel blind source separation are difficult to solve this new alias mode. This dissertation focuses on blind separation based characteristic domain method for permuted alias image.Information integrity of permuted alias image is losed for it is composed of different segement of source images with unoverlapping permution. Various sources of images make the permuted alias image diversity. At the same time, the position size and number of permuting image are unknown. The key of detecting and seperating permuting images is to find characteristic difference, which can be shown when permuted alias image is projected on a characteristic domain. This dissertation proposes methods of blind separation in various characteristic domain for various type of permuted alias images. The contributions of the dissertation are summarized as follows.1. A blind separation method based sparse decomposing is presented according to a type of permuted alias image, when morphological diversity exists betweeen permuted and permuting region of image, which can be expressed sparsity diversity decomposed on different dictionaries. Sparsity representation model of permuted alias image is obtained by its mathematical model. On the base of characteristics of permuting and permuted region image, Contourlet and Local DCT base dictionary are chosen as characteristic domain. Sparse decomposed iterately on two different charactersitic domains, getting sparsity representaion of different morphological images by block coordinate relaxation method, the permuted image can be seperated from the permuted alias image. Results show that, as for a permuted image comprising piecewise smoooth part and texture part, our algorithm can better separate texture image from the permuted image, not affected by size, location, number and types of texture image.2. Different region of permuted alias image have noise difference, for permuting and permuted images come from different source images and they usually be infected by noise in the process of acquisition, transmission and processing. An algorithm about permuted alias image blind separation based noise detection is proposed according to a class of permuted alias image with noise difference. Learning dictionary is chosen as characteristic domain, for it has high adaptability and its sparsity representation has high sparsity degrees than that of parameter dictionary. Permuted alias image is denoised by getting its sparsity representation with K-SVD restrained by nonzeros number dictionary learning algorithm with itself smapling. Size and location of permuting region is found out by detecting the subtraction image, which is defined as difference between the denoised permuted alias image and original permuted alias image. The permuting region is optimized by implementing image morphological operation and is separated from the permuted alias image by setting threshold. The results show that the region noised can be effectively denoised only with two constant nonzeros number and permuting image can be separated efficiently from the permuted alias image, not affected by size, location, number of permuting image and noise level on permuting image.3. A blind separation algorithm based on four-direction-difference is proposed for a type of permuted alias image with motion blur difference. Firstly, space domain is chosen as characteristic domain after analysing characteristics of motion blur image. Four differential images in four directions are obtained by four direction operating of permuted alias image. A binary image can be obtained by calculating variance value of each differencd image subblock and thresholding it, to detect rough location of permuting image. Permuting images can be seperated from permuted alias image by multiplying binary image which is optimized with image morphological operation. The results show that blurry images can be effectively separated from the permuted alias image without respect to the location, size and number of motion blur image and direction of motion blur.4. A blind separation algorithm based on differential evolution is proposed for a type of permuted alias image with gaussian blur difference. Firstly, space domain is chosen as characteristic domain after analysing characteristics of gaussian blur image. Differential image of permuted alias image is divided into sub-blocks which is randomly assigned a threshold and all the thresholds formed a threshold vector. The differential evolution is then performed to obtain the optimal threshold vector for thresholding the differential image into binary one. The permuting image could be separated by permuted alias image multiplid by the binary image optimized with image morphology operating. Experimental results show that the proposed approach could effectively separate the permuting image from the permuted alias image without respect to the location, size and nemuber of permuting image.更多还原