基于深度学习的四元数小波彩色图像质量评价及其应用

项目名称： 基于深度学习的四元数小波彩色图像质量评价及其应用

项目编号： No.61461043

项目类型： 地区科学基金项目

立项/批准年度： 2015

项目学科： 无线电电子学、电信技术

项目作者： 刘国军

作者单位： 宁夏大学

项目金额： 45万元

中文摘要： 图像质量评价(IQA)是当前图像工程一个非常活跃的研究领域，其本质是寻求符合于人的主观视觉的客观IQA方法。颜色信息作为十分重要的视觉特征之一，对于构建彩色IQA方法具有举足轻重的意义。已有的彩色IQA方法大多采用色度和亮度分离策略，过于简化了它们之间复杂的多层非线性相关性。本项目拟采用系统研究的理念，利用四元数及其相关理论设计深度学习框架下的全参考型和无参考型彩色IQA模型和算法。主要内容包括：1、利用纯四元数表示彩色图像，构建基于深度学习的四元数小波(QW)彩色IQA模型和算法，这也是本项目的研究重点和亮点。2、利用Hilbert变换构造灰度图像的四元数表示，利用四元数矩阵的奇异值分解、主成分分析、QW等理论，构建深度学习的灰度图像IQA方法；3、将新的彩色IQA方法应用于彩色图像滤波、分类等问题。本项目有望获得物理意义清晰、数学理论可靠、计算简单有效、适合于不同失真类型的IQA方法。

中文关键词： 彩色图像；四元数小波；深度学习；图像质量；客观评价

英文摘要： Image quality assessment (IQA), the purpose of which is to seek an objective IQA method to be consistent with the subjective vision, is one of most active areas of recent researches in image engineering. As a type of very important feature of the human visual system(HVS), the color information plays its own irreplaceable role in a color IQA method. The research results of neuroscience show that knowledge of these color-luminance relationships is built into the machinery of the human visual system. However，most of the existing color IQA methods perform strategy separating chrominance with luminance components, which simplifies the non-linear multilayer correlation between image visual features. In this project, we adopt a quaternion framework since it offers the scope to process color imges holistically, rather than as separate color space components, and thereby handeles the coupling between the color channels naturally. More specifically, we will design both full-reference and no-reference IQA models and algorithms based quaternion and some related theories under the framework of deep learning. Our main work includes three aspects as follows. First, after representing a color image with a pure quaternion, we will construct a deep convolution network architecture with quaternion wavelets(QW) transform to design color models and algorithm, which is both an important issue and a highlight of this project. Second, after representing a gray image with a quaternion based two dimensional Hilbert transform, we will propose IQA method based on a deep learning framework by emploring some theories of quaternion, such as quaternion matrix singular value decomposition, principal component analysis, QW, et al.. Finally, the proposed color IQA methods will be used for color image processing and pattern recognition, such as filtering and classification, et al.. In a word, we hope to obtain IQA methods with a clear physical meaning, a reliable mathematical theory, a simple and efficient computation, as well as a general application for different types of distortion.

英文关键词： Color image;Quaternion wavelets;Deep learning;Image quality;Objective assessment