【作者】 程光权；

【导师】 成礼智；

【作者基本信息】 国防科学技术大学 ， 计算数学， 2010， 博士

【摘要】 随着科学技术的发展,以及人们需求的日益提高,从一维信号处理中发展起来的经典图像处理算法,已越来越难满足人们对高质量图像处理的需求。小波等经典图像处理方法忽略了高维数据的本征几何结构特征,并不是适于图像数据结构的视觉最优图像处理方法。因此,为了改善各类图像处理算法的效果,必须从图像数据的本身结构特点出发,根据人眼视觉系统特性,结合实际应用背景需求,设计真正适合于图像数据特征的图像处理算法。图像数据的离散属性,以及结构的复杂特性,决定了在数字图像处理过程中,建立符合视觉感知特点的适于应用背景的精确模型的困难性。本文在深入研究图像数据视觉感知特点的基础上,针对传统图像压缩、分辨率增强以及质量评价算法设计中的不足,对算法设计中的一些关键问题做出了深入研究。本文的主要工作和创新包括以下几个方面:1.针对基于小波变换图像压缩方法的不足,结合图像数据几何结构特征,提出了基于边缘导向的正交小波变换图像压缩方法。该方法在继承经典小波变换优点的基础上,能够充分理解图像数据的方向奇异结构特性,有效地利用图像数据空间不均匀的特性,提高图像压缩效率的同时,有效保护图像数据中人眼视觉感兴趣的几何奇异特征。同时,将方法根据SAR图像数据特点加以改进,拓展应用到SAR图像压缩中。2.针对传统插值方法的不足,提出了基于小波变换的边缘保持方向自适应图像插值方法。该方法通过改进双线性插值方法,自适应调整插值核函数,有效地保护了图像的边缘特征。同时结合小波变换的多分辨表示性能,有效地提高了插值图像的高频信息,并进行相关后处理,增强了图像的视觉效果。与传统方法比较,试验结果的主客观质量都得到了提高。3.在研究二分树复小波变换系数几何先验信息的基础上,建立了基于复小波变换的超分辨图像重建模型。该方法利用二分树复小波变换具有近似平移不变和灵活的方向选择性,实现图像的高效稀疏表示。同时,根据复小波变换系数模值和相位信息在边缘处的几何约束条件,结合超分辨重建问题,从而在复小波变换域建立一种新型的超分辨重建模型。最后,利用分裂Bregman方法实现模型的有效优化求解,得到高质量的超分辨率图像。4.针对传统质量评价方法的缺陷,根据视觉感知图像数据的特点,提出了基于几何结构失真模型的完全参考图像质量评价方法。该方法根据图像数据中引起视觉敏感的方向失真、幅度失真和锐度失真,建立了几何结构失真模型,物理意义明确,计算复杂度较低,符合人眼视觉感知特点,试验结果与主观预测结果具有很好的一致性。同时,利用小波变换与人眼视觉系统的多通道特性相匹配的特点,建立基于小波变换的几何结构失真模型的质量评价方法,试验结果验证了方法的有效性。5.根据自然图像统计先验信息,提出了基于边缘特征统计的部分参考型图像质量评价方法。图像边缘信息在人眼感知图像质量过程中占据着十分重要的地位,而自然图像的边缘统计分布符合一定的先验统计规律,该方法通过度量这种统计分布特征的变化程度预测图像质量,仿真试验对标准图像库中所有失真类型数据都得到较好的预测结果。总之,本文从图像数据结构特征出发,结合人眼视觉感知特性,解决基于方向小波图像处理算法与几何特征保持质量评价方法设计中的一些关键问题,获得了更加符合人眼视觉系统特性的试验结果。更多还原

【Abstract】 With the development of technology and the increase of requirement, it is becoming more and more difficult to obtain visually satisfactory result by conventional image processing algorithms. The limitations of commonly used separable extensions from one-dimensional transforms for images, such as discrete wavelet transform (DWT), are well known; e.g. these separable transforms cannot take advantage of intrinsic geometrical structure in high dimensional signals and are not the optimal algorithms for image. To improve the performance of image processing algorithms, it is necessary to explore the geometry in image and the characteristic of human visual system which are desirable features for many consumers and practical applications. That is to say, the image processing algorithms should be directly driven by the structure of image data.The performance of image processing algorithms closely relies on the accuracy of employed models to characterize the image. However, it is difficult to construct an accurate image model in according to visual prediction and practical applications. The main challenge in exploring geometry in images comes from the discrete nature of image data and complexity of image structures. In this thesis, we study the characteristic of image data and human visual system and solve some crucial issues in algorithms to overcome the disadvantage of traditional image compression, image resolution enhancement, and image quality assessment. The main contributions of this thesis could be summarized as follows:1. In order to overcome the disadvantage of conventional wavelet based image compression, we propose an edge-directed orthogonal wavelet transform which is driven from the geometric structural feature of image. The proposed method inherits the advantage of wavelet and explores the directional features of images. Different schemes are implemented based on the property of image blocks to reduce the computational complexity. Experiments show that the new method can protect effectively the geometric feature which plays an important role in visual perception. Meanwhile, the extension of this method to wavelet packets for SAR image compression is straightforward.2. The blur and jaggy of image details or edges are inevitable during conventional image interpolation. In order to obtain interpolated images with better quality, we propose wavelet based edge-preserving direction adaptive image interpolation method. We apply the improved bilinear interpolation method with adaptive direction to interpolated image. Wavelet is implemented to provide more high frequency information, and post-processing is applied to improve the visual quality of interpolated images. The experimental results show that our method can achieve interpolated image with high quality, both subjectively and objectively. 3. A novel single image super-resolution reconstruction algorithm is proposed based on the geometrical model on the phase and amplitude of dual-tree complex wavelet coefficients of the image. The dual-tree complex wavelet has the properties of approximate shift-invariance and flexible directionality, and can achieve sparser image representation compared with standard wavelet. The appropriate geometric regularization is designed based on the priors of the amplitude and phase of complex wavelet coefficients for super resolution image reconstruction. Then, Split Bregman iteration is utilized in our proposed approach for optimization to gain high quality super resolution image.4. Inspired by the researches of quality prediction of human visual system and the intrinsically geometric structural features of natural images, a novel geometric structural distortion model based full reference image quality assessment method is proposed to overcome the deficiencies in traditional methods. Basically there are three components in our measurement to characterize the geometric structural distortion: direction, magnitude and sharpness. The proposed measurement fits the physical observations for various image distortions and has relatively low computational complexity. The experimental results on image database show that the performance of our method is consistent with the subjective assessment of human beings. Meanwhile, wavelet based geometric structural distortion is proposed in according with perceptual property of human eye, where wavelet transform is used because it matches well the multi-channel model of HVS. The experimental results demonstrate the advantage of proposed model.5. A novel reduced reference image quality assessment based on natural image statistical prior in gradient domain is proposed. The research in human visual system shows that edge information plays an important role in visual perception. On the other hand, it obeys a specify distribution for natural image, where some statistical features of reference image are extracted and sent to receiver side. The distortion measure for distorted image is defined with comparison of these features. The experimental resuts on standard image database shows that proposed method is general purposed for all distortion types.In summary, following the characteristic of image data and human visual perception, this thesis provides several systemic researches about the problem of direction adaptive wavelet based image processing algorithms and geometric structural features preserving image quality assessment.更多还原