【作者】 张骥祥；

【导师】 戴居丰；

【作者基本信息】 天津大学 ， 通信与信息系统， 2007， 博士

【摘要】 本论文根据小波变换和马尔可夫随机场理论,针对图像处理技术中的图像去噪、图像分割和纹理合成三个重要课题进行了深入研究,主要工作及贡献如下:1.提出了一种新的基于小波域隐马尔可夫树模型的图像去噪算法。该算法采用三态隐马尔可夫树模型或分级隐马尔可夫树模型对图像进行多尺度统计建模,降低了计算复杂度,提高了模型参数的估计精度。对加性高斯白噪声进行图像去噪仿真实验,实验结果表明本文提出的图像去噪算法在峰值信噪比和主观视觉效果方面均优于传统图像去噪算法。2.提出了一种基于小波域自适应上下文结构的多尺度图像分割算法。该算法为了减小计算量,采用智能初始化和半树HMT模型参数加权训练算法,得到了可靠的初始分割;为了获得较好的区域一致性和边缘准确性,在进行尺度间融合时,采用自适应的上下文结构分别应用于图像纹理均质区域和图像纹理边缘,保证了图像大致轮廓的准确性和可靠性,提高了分割后图像纹理边缘的精确度。仿真实验结果表明本文提出的图像分割方法对合成纹理图像和自然界图像均有很好的分割效果。3.提出了一种新的基于马尔可夫随机场的纹理合成方法,该算法首先利用纹理图像及其子图像的统计特性的相似性,自动选取最佳邻域的大小,然后采用L邻域匹配快速搜索算法加速纹理合成。仿真实验结果表明,本文提出的纹理合成算法和传统算法相比,不仅减少了合成时间,而且获得了比较好的纹理图像合成效果。更多还原

【Abstract】 In this dissertation, we focus on three important topics of image processing based on wavelet transform and Markov random field theory, which are image denoising, image segmentation, and texture synthesis. The main contributions are as follows:We develop a new approach to the wavelet-domain hidden Markov tree model for removing noise from the images. In order to undertake the accuracy of the model parameters and reduce computing expense, we use 3-state hidden Markov tree model(3SHMT) or hierarchical hidden Markov tree model(HHMT) for multiscale statistical image modeling. We demonstrate through simulations with images contaminated by additive white Gaussian noise that the performance of this method substantially surpasses that of previously published methods, both visually and in terms of PSNR.A new image segmentation algorithm based wavelet, referred to as joint adaptive context and multiscale segmentation(JACMS) is developed. Towards achieving lower computational complexity, we propose an intelligent parameter initialization and half tree HMT model weighting training algorithm, when applied to image segmentation, this technique provides a reliable initial segmentation. In order to achieve higher accuracies of both texture classification and boundary localization during the interscale fusion, we develop adaptive context structures that are applied to homogeneous regions or/and texture boundaries, respectively. Experiments demonstrate that the proposed algorithms yield excellent segmentation results on both synthetic and real world data examples.A new MRF-based texture synthesis algorithm is proposed. First, the optimal neighbor size is automatically chosen using statistic characteristics similarity between the texture image and its subimage, then the L neighbor matching fast search algorithm is introduced to accelerate texture synthesis. Our texture synthesis methods can produce better results in little time than previous methods.更多还原