【作者】 李兵；

【导师】 须德；

【作者基本信息】 北京交通大学 ， 计算机应用技术， 2009， 博士

【摘要】 知觉恒常性作为人类视觉系统最基本和最重要的功能之一,对人类正确稳定地感知世界起着极其重要的作用。为了使计算机视觉系统具有类似的感知功能,知觉恒常性计算成为了计算机视觉领域中一个重要的研究课题。论文以知觉恒常性中的颜色恒常性作为研究对象,从无监督的颜色恒常性计算、有监督的颜色恒常性计算、颜色恒常性算法的融合、以及颜色不变性描述四个方面展开了颜色恒常性计算的研究:在无监督的颜色恒常性计算上,针对目前GSI(Grey Surface Identification)算法依赖于相机参数的缺点,提出了一种基于灰色表面的颜色恒常性算法(ColorConstancy using Achromatic Surface,CCAS)。该算法利用迭代的方法来进行灰色表面提取,从而避免了需要获取相机的成像参数。利用CCAS算法得到的灰色表面来估计图像的光照颜色,可以很大地提高光照估计的准确度。在有监督的颜色恒常性计算上,为了克服基于支持向量回归的颜色恒常性算法的单一输出、参数调节繁琐、以及学习速度缓慢的缺点,引入了一种全新的单隐藏层前向神经网络的学习算法ELM(Extreme Learning Machine)算法,提出了一种基于ELM的颜色恒常性算法。此外,为了克服传统的二值化的色度直方图特征维数过高的缺点,论文基于Grey Edge算法框架提取了一种新的低维高效的图像特征作为基于ELM的颜色恒常性算法的输入向量。在颜色恒常性算法的融合上,利用威布尔(Weibull)分布参数的图像纹理描述方法,综合考虑图像的全局纹理特征和局部纹理特征的基础上,提出了一种基于纹理相似性的自然图像的颜色恒常性计算算法。该算法利用Grey Edge算法框架作为颜色恒常性算法的产生器,根据图像的纹理特征,为其选择一个最优的颜色恒常性算法或算法组合。在颜色不变性描述上,基于鲁棒的Diagonal-offset反射模型,定义了两个新的颜色空间。在两个新的颜色空间上,引入了不变矩理论,提出一个基于不变矩的颜色不变性描述子的计算框架。根据不同的颜色空间选择,该框架可以产生:原始图像的颜色不变性描述子和边缘图像的颜色不变性描述子。而这两个描述子的融合又构成了融合的颜色不变性描述子。这些颜色不变性描述子不仅具有对图像光照变化的鲁棒性,而且还具有对图像几何变化以及图像模糊的鲁棒性。更多还原

【Abstract】 As one of the most basic and important visual functions, perceptual constancy plays a key role for the human beings stably perceiving the real world. In order to make computer vision system have the similar visual perception function, perceptual constancy computation has become one of the most important research topic in the computer vision field. This dissertation focuses on the color constancy computation, and studies this topic from four aspects: unsupervised color constancy, supervised color constancy, color constancy algorithms fusion and color invariant.For the unsupervised color constancy computation, the recently proposed color constancy algorithm GSI (Grey Surface Identification) is a camera-dependent solution and suffers from camera calibration requirement. In order to avoid calibrating a camera, we propose an iterative algorithm, which is called color constancy using achromatic surface(CCAS). The achromatic surfaces identified by the CCAS algorithm can be used to estimate illumination of an image, which can significantly improve accuracy of illumination estimation.For the supervised color constancy computation, by introducing a novel machine learning algorithm Extreme Learning Machine (ELM), we propose a new color constancy algorithm named color constancy using ELM, which can overcome the single output, complex parameter tuning and low training speed of the SVR based color constancy algorithm. In addition, discarding the traditional high dimensional binarized chromaticity histogram feature, a new low dimensional and effective feature vector is extracted based on the Grey Edge framework, which is used as the input for the ELM based algorithm.For the color constancy algorithms fusion, based on the image texture description with Weibull parameterization, we propose a new color constancy fusion scheme called color constancy based on texture similarity for natural images. By using the Grey Edge framework as the algorithm generator, the proposed fusing algorithm can select the best color constancy method (or best combination of the methods) for the certain image, according to its both global texture feature and local texuture feature.For the color invariant description, two new color spaces are defined based on the robust Diagonal-offset reflectance model. In the two color spaces, a color invariant descriptor computation framework is proposed, by introducing moment invariants theory. According to different color space selection, the descriptor computation framework can generate original image color invariant descriptor and edge image color invariant descriptor. The combination thereof can serve as combinational color invariant descriptor. All these descriptors are not only robust to illumination change, but also robust to affine transformation and image blurring.更多还原