【作者】 向遥；

【导师】 邹北骥；

【作者基本信息】 中南大学 ， 计算机应用技术， 2011， 博士

【摘要】 大量可以被人类和其它动物毫不费力完成的视觉任务,如在照片中识别朋友的面容,理解一个从未见过的物体是飞机还是轮船,用当前的机器视觉或图像处理系统却难以实现。人类及其他灵长类动物的视觉系统,能够实时而充分地感知外界的环境信息,并迅速地做出判断,无论从哪方面比较和衡量,都优于目前最好的机器视觉和图像处理系统。因此,探究人或灵长类动物视觉系统对外界场景进行感知和信息处理的过程,并以其为参照来设计能完成特定视觉任务的图像处理算法,即模拟人视觉感知的图像处理方法,已成为当前用计算机完成视觉任务的新的研究思路和热点。遵循这种研究思想,本文针对颜色迁移、彩色图像去色和用于全景拼接的图像排序等图像处理问题中所涉及的视觉感知现象和模型进行分析,进而提出能解决这三类问题的新的基于视觉感知的处理方法。具体工作如下：(1)颜色迁移。选择在(?)αβ颜色空间中完成颜色迁移的处理,该颜色空间是利用大量的自然界图像来进行驱动人的视觉感知的实验而得到的一个统计上各分量间相互独立的空间,其灰度和色彩分量分离,近似模拟视网膜->侧膝体中神经元对可见光信号的刺激响应和处理过程。针对已有颜色迁移算法仍存在的问题,提出了适用于复杂场景的多源图像颜色迁移算法；并从色彩相似度和结构相似度两方面对迁移合成图像质量进行衡量,提出了适用于颜色迁移算法的客观评价指标；大量实验结果验证了所提算法的正确性和评价指标的有效性。(2)彩色图像的去色。提出了两种彩色图像到灰度图像的转换算法：一种是采用数理统计分析的方法,尝试从图像中寻找色彩的差异值,并以此对物理明度分量进行增强；另一种是从彩色图像去色问题所涉及的视觉感知问题的角度进行分析。利用Helmholtz-Kohlrausch色貌效应所定义的主观亮度指标来确定彩色图像的物理明度和饱和度的顺序关系,并将该主观亮度指标与色调作为视觉注意选择算法PFT(Phase-spectrum of Fourier Transform)的两个输入变量,计算后获得图像的色彩显著性图。由于该色彩显著性图综合反映了原彩色图像的明度、饱和度和色调的显著性差异,最终得到的灰度图像更符合人的视觉感知。(3)用于全景拼接的图像序列排序。提出了两种全景图像序列自动排序算法：一种采用传统的图像处理方法,在梯度域计算得到图像的相似曲线,并以此作为特征来实现对图像序列的排序；另一种从视觉感知的层面出发,分析人在判断图像序列拍摄顺序时的处理方法,利用视觉注意力选择算法PQFT(Phase spectrum of Quaternion Fourier Transform)在频域中计算两两图像之间相位谱的偏移量,并以此为特征,得到一种新的图像序列排序算法。实验结果验证了这两种算法的有效性。更多还原

【Abstract】 A number of visual tasks which can be effortlessly achieved by human and other animals have still remained seemingly intractable for computers。These tasks include recognizing the face of a friend in a photograph, understanding whether a new, never-seen object is a car or a stream boat. The visual system of human and other primate animals are complicated and accurate, and be able to sense the external environment immediately and then make judgment rapidly. No matter measure from which side, human vision system is better than the current best machine vision and image processing system. Exploring the sensing and information processing mechanism of the primate visual system, and using them as reference to design image processing algorithms provide new research approaches to accomplish visual tasks by computers; Such approaches are classified as image processing methods based on visual perception. In this paper, three image processing cased, which are color transfer, color image to grayscale conversion and image sorting for panorama image mosaic, are chosen as research cases, and new solutions to these cases based on visual perception phenomenons are studied. Our research work consists of three parts as following:In the first part, color transfer is studied. We choose to perform our algorithm under (?)αβcolor space, which is proposed by simulating the photostimulation signal processing by retinal ganlion cells and lateral geniculate nucleus. The data of (?)αβcolor space are expressed in an orthogonal decorrelation space robustly producing three principal axes, one is corresponding to simple changes in luminance and the other two that are reminiscent of the blue-yellow and red-green chromatic-opponent mechanisms found in the primate visual system. By analyzing the unsolved problems in current color transfer algorithms, we propose a new solution by using multi-source images as color references, which is applicable for re-colorizing images with complex scenes. Also, an objective evaluation index, considering color similarity between the result image and source image and structure similarity between the result image and object image, is proposed to measure color transfer algorithms. Experimental results verify the accuracy of our proposed algorithm and the effectiveness of the evaluation index.In the second part, we studied the problem of converting color images to grayscale ones, and proposed two new algorithms. One is based on mathematical and statistical analysis, aimed to find the differences between colors and map them to the physical luminance channels. The other algorithm is proposed by analyzing the visual perception phenomenon involved in image de-colorization. PFT (Phase-spectrum of Fourier Transform) visual attention selection algorithm is used to obtain the color saliency map of the color image, whose two input variables are carefully designed. One variable is a chromatic lightness term based on Helmholtz-Kohlrausch effect that corrects physical luminance based on the color’s saturation component, and the other is the hue component of the color image. Hence, the saliency map synthetically reflects the differences of luminance, saturation and hue components of the color image, which makes the result grayscale image generated by the proposed algorithm looks more consistent with the visual perception.In the third part, image sorting for panorama image mosaic is studied, and two algorithms to solve this problem are proposed. The first solution is considered as a traditional image processing method:we define similar curves in the image’s gradient domain, and take the curves as image characters to perform the sorting method. The second algorithm refers to the process of human judging the image sequences, and uses the visual attention selecting strategy-PQFT (Phase spectrum of Quaternion Fourier Transform) to perform sorting. In this case, images are converted into frequency domain pair wise, and the translation distances between any two images are calculated based on their phase spectrum. Experimental results demonstrate the effectiveness of the two proposed algorithm.更多还原