【作者】 沈会良；

【导师】 李志能；

【作者基本信息】 浙江大学 ， 电子科学与技术， 2002， 博士

【摘要】 图像理解是研究用计算机系统解释图像，实现类似人类视觉系统理解外部世界的一门科学，它主要讨论从图像中获取信息和解释信息两个方面。自Marr于二十世纪八十年代提出视觉理论以来，图像理解学科有了很大发展，但还存在着较大的困难，主要体现在如何设计能适应复杂场景解释的算法及实时实现等方面。本文重点研究了中低层图像理解算法，包括图像恢复、图像分割、特征提取、纹理图像颜色映射等方面。 在图像采集过程中，由于实际环境中成像条件的限制或物体表面特性的影响，造成图像的背景光照不均匀，影响到后续图像处理步骤的效果。针对这一问题，本文提出了基于背景变化拟合的图像恢复和增强算法，取得较好的效果。该算法首先从图像中提取空间位置均匀分布的背景样点，然后作多项式拟合得到背景光照的分布特性，开通过调节对比度因子，达到恢复和增强图像的目的。 在图像特征提取方面，首先研究了小波的多分辨率特性，并提出了基于小波分析特征检测的二值字符图像识别算法和特定物体——汽车牌照分割算法。在字符识别算法中，首先提取字符图像的线性矩，并对线性矩和字符图像分别作一维和二维小波变换，然后将多个尺度上小波系数的统计变量作为特征识别，达到了很高的正确率。在牌照图像分割算法中，克服了以往基于竖向纹理方法容易受到干扰区域影响的缺陷，能够更加快速准确地完成灰度图像中车牌的定位。 在彩色图像分割方面，分别提出了基于直方图层次分析和基于差值分水岭的分割方法。直方图层次分析法在CIELAB颜色空间进行，通过对三个一维直方图的循环分析和区域合并达到图像分割的目的。差值分水岭方法考虑了象素的空间位置分布，且根据同一区域中象素间差异小于不同区域象素间差异这一相对较弱的假设，克服了以往梯度分水岭方法不能有效分割区域间较宽过渡带的缺陷，达到较好的分割效果。 在研究彩色纹理图像中RGB各通道空间分布特性的基础上，提出了纹理图像的颜色精确映射算法。该算法针对原始图像的不同有两种模式，即灰度—彩色映射(GCM)模式和彩色—彩色映射模式(CCM)模式。在GCM模式中，只有亮度信号的空间分布是已知的，所以算法需要从单通道信号中推导出彩色图像的三通道信息，而在CCM模式中，由于初始图像是彩色的，完成从三通道到三通道的信息映射。模拟计算值(客观)和视觉直观(主观)评估实验证明了所提出算法在颜色映射中的有效性，为图像压缩、通信和计算机辅助设计提供了有益思路。 最后，研究并实现了一个实际图像分析系统——汽车牌照识别系统。详细讨论了该系统的中图像处理的各个步骤，包括图像的分割、二值化、牌照上各字符的切分和识别，并在分析大量实际图像的基础上，提出了有效的算法。系统在实际运行环境中达到96％以上的正确识别率，已通过科技成果鉴定，达到国内领先水平，并得到了大量推广应用，获得了较大的经济效益。更多还原

【Abstract】 Image understanding is a subject about interpreting image using computers by imitating the biological systems of human beings. There are two parts involved, which are capturing the image content and establishing the relationship among objects. Since the vision theory proposed by D.Marr in the 1980s, researchers make great progress in the area of image understanding. However, it encounters two difficulties, which are (1) designing adaptive algorithms capable of effectively understanding the most common and complicated scene, and (2) completing vision task effectively and understanding images in real time. This thesis emphasizes on the study of middle and low level algorithm in image understanding, including the section of image restoration, image segmentation, character extraction, color mapping on texture images, and etc.Due to the limitation of imaging condition and the property of the object surface, there are sometimes defects of nonuniform illumination in the captured images. To correct these defects, an algorithm of restoration and enhancement based on background function fitting is developed. In the algorithm, the uniformly spaced background sample pixels are automatically selected, and then they are used in polynomial fitting for background calculation. By adjusting the factors according to the statistical parameters, image enhancement can be achieved.In the part of character extraction, the property of multi-resolution in wavelet is carefully studied and applied in two algorithms, one is character recognition, and the other is license plate extraction. In the algorithm of character recognition, the wavelet transform is firstly applied on the linear moments and the image itself, and the statistical variable is then calculated as the input of the three-lays back-propagation neural network for recognition. For the algorithm of license plate extraction, thanks to the excellent performace of wavelet analysis in edge extraction, the locating of the plate region is more accurate and robust, compared with the former algorithm utilizing vertical edge texture of the plate.Two algorithms-hierarchical histogram analysis and watershed of difference (WOD)-are presented in the area of color image segmentation. In the former algorithm, the segmentation is reached by analyzing the three one-dimensional histograms in CIELAB color space recursively, followed by the operation of region merging. In the latter one, the spatial location of pixels is considered to ensure the compactness of the regions. As there are always wide boundaries between different regions, within which the values of the pixels vary gradually from one region to another, it is difficult to classify the pixels in that boundary region. According to the homogeneous character of region, we draw a hypothesis that the difference between pixels in the same region issmaller than that in different regions. Based on this hypothesis, the algorithm called watershed of difference performs better in segmentation than the traditional algorithm of watershed of gradient (WOG).Based on the careful investigation of the inter-relationships among spatial distribution in the red, green, and blue channels of the color textural images, a computational model is developed. Because of the difference in original image information, the computational model has two modes: gray-to-color-mapping (GCM) and color-to-color-mapping (CCM). In GCM mode, only the spatial distribution of luminance is available, the algorithm needs to deduce three-dimensional information of color image from one-dimensional information; whereas in CCM mode, the information in each of the three channels is available. The model proposed in this paper gives excellent mapping results both numerically and according to the psychophysical experiment. The algorithm may offer a possible solution in the area of image compression, communication, and CAD system.In the end, a practical image analysis system-license plate recognition system-is proposed. The paper discussed the four major steps in the realization更多还原