基于数据驱动的颜色恒常性计算-color constancy calculation based on data driving.docxVIP

摘要作为最重要的视觉特征之一，颜色被广泛应用于计算机视觉的诸多领域，如 图像分割、对象跟踪、物体识别、图像检索等。但是，颜色很容易受到光照变化 的影响从而使得基于颜色的算法对光照不具有鲁棒性，因此迫切需要解决在不同 光源条件下物体表面颜色描述的一致性问题。本文围绕颜色的恒常性问题展开研 究，旨在解决如何将未知光源下的颜色（信号）转变到在标准的已知光源（通常 为白色）下的颜色值。考虑到色域数据都来源于同一光源，因此，它们之间应该存在某种内在的联 系，基于此，本文试图发现出此种联系并将其作为先验知识来求解颜色恒常性问 题。因此，本文的主要工作包括：1、提出了一种新的数据拟合方法——支持向量拟合来发掘色域数据之间的内 在联系。本文通过形式化推导指出同一光源下的观测值之间存在着某种内在的联 系，这种联系表现为隐藏在光源色域数据中的复杂的非线性关系，本文提出的支 持向量拟合方法，能够将数据从低维空间隐式地映射到高维空间，并使其在高维 空间具有超平面结构，从而可以用超平面的结构系数来隐式地描述该光源，并将 此作为解决颜色恒常性问题的先验知识；2、提出了基于数据驱动的颜色恒常性算法。首先对标准光源的色域进行训练， 通过支持向量拟合来形式化描述标准光源，即将光源估计问题转化为超平面参数 估计问题，最终将颜色恒常性问题转换为最优化问题，即通过最优化模型找出映 射矩阵，通过该映射矩阵使得未知光源下的图像色域在调整后能够在高维空间中 最好地符合标准光源所对应的超平面；3、利用国际上通用的测试图像集，对本文提出的方法及其他经典算法进行了 测试和分析比较。实验表明，本文算法的平均角度误差和中值角度误差都小于现 有的颜色恒常性算法，在各种光照条件下都能取得较好的效果。主题词：颜色恒常性；色域；数据驱动；支持向量拟合ABSTRACTAs one of the most important feature, color is generally used in many computer vision tasks such as image segmentation, object tracking, object recognition and image retrieval. However, the variant of the illuminant can affect the color of the object dramatically which cause the algorithms depend on the color will not robust to the light, therefore, it is urgent to solve the problem that the color of the object surface under different illuminant conditions needed to appeared the same. This thesis focuses on the computational color constancy and solves that problem by transforming the image taken under unknown illuminant to that under the canonical, known light.Considering the gamut data have the same source, therefore, there may be some unknown inner relationship among them. Therefore, the task of this thesis is to attempt to uncover this relationship which could be treated as a priori, thus, the work of this thesis include as follows:We propose a new data fitting method—support vector fitting (SVF) which is used to exploit the inner relationships among the gamut data. This thesis concludes there exists some relationship among observed data by formally deriving. This relationship implicitly represents as the complica