【作者】 许欣；

【导师】 夏德深；

【作者基本信息】 南京理工大学 ， 模式识别与智能系统， 2010， 博士

【摘要】 图像增强是指按特定的需要采用特定方法突出图像中的某些信息,同时削弱或去除无关信息,或将原图转换成一种更适合人或机器进行分析处理的形式的图像处理方法,其设计与其应用的目的密切相关.近年来,随着消费型和专业型数码相机的日益普及,海量的图像数据正在被产生.但由于场景条件的影响,很多在高动态范围场景、昏暗环境或特殊光线条件下拍摄的图像视觉效果不佳,需要进行后期增强处理调整动态范围或提取一致色感才能满足显示和印刷的要求.人类视觉系统具有强烈的全局和区域的自适应性和非线性,在多种不同的光照条件下都能清晰地辨识细节,具有电子设备所不可比拟的优势.因此,很多图像增强方法在设计时考虑描述和模仿人类视觉系统的特性,以期获得符合人类视觉系统特性的增强效果.本学位论文对实际应用中常见的几类改善图像视觉效果的增强理论和技术进行了深入研究,对它们在实际应用中遇到的问题和存在的缺点进行了分析,提出了几种图像增强的改进方法.论文主要包括以下几部分内容：(1)提出了一种快速Retinex图像增强方法,模拟了人类视觉系统的全局和局部自适应性.由于假设场景中光照是平缓变化的,传统的中心／环绕Retinex图像增强方法在处理高动态范围图像时易在明暗对比强烈处产生光晕现象.提出的方法在光照估计步骤中对原图进行Mean Shift滤波,用于其后光照影响的消除.由于自适应滤波的边缘保持能力,增强结果的光晕现象得到了有效的消除.实验结果表明,本方法能够有效压缩图像的高动态范围,克服光照不均的影响.由于使用了Mean Shift加速算法,本文方法运行速度快于此前提出的方法.(2)提出了一种改进的自动颜色均衡化方法,用于图像对比度的增强.首先通过考虑图像中颜色或灰度的空间分布对图像进行了局部自适应滤波,其中使用了改进的相对亮度表观函数,减小了对最小梯度的拉伸.而后对图像进行动态范围调整以得到最终结果.在对彩色图像处理时,通过保持RGB通道间的比例保持颜色不失真.实验结果表明,方法可有效增强图像的对比度,且不会引入明显噪声.(3)提出了一种改进的结合视觉感知特性的变分框架下的彩色图像增强方法.设计了一种考虑图像局部特征的区域自适应参数用来控制局部对比度增强的程度,采用梯度下降法求解能量泛函的极小值.实验结果表明,方法能有效增强彩色图像的对比度,图像较暗处和平坦处的细节得到了有效的改善.(4)提出了一种结合视觉特性的梯度域图像增强方法.常见的图像对比度增强方法在压缩图像整体动态范围的同时有效增强图像较暗处的细节,但图像原较亮部分的细节往往得不到增强,甚至被削弱.针对人类视觉感知的特性,提出了相对梯度的概念,首先在梯度域对图像原较亮处的梯度进行更大的拉伸,然后在最小二乘意义下重建出增强后的结果图像.实验结果表明,方法可有效压缩图像的整体动态范围,同时原图像中较暗和较亮处的细节都得到了有效增强或保持.(5)针对同一图像采用不同增强方法处理的结果之间可存在互补优缺点的特点,提出了采用梯度域融合的方法改善图像视觉效果的增强方法.首先将待融合各图像的结构张量按一定比例进行融合,在权重的设计中考虑了各通道图像的局部对比度.之后求出目标梯度场,其结构张量在Frobenius范数意义下逼近前述融合后得到的结构张量.另外在融合模型中加入了与感知特性相符的增强能量项,进一步改善增强结果的视觉效果.方法还可应用于相同场景采用不同对焦距离或不同曝光时间所拍摄照片之间,以及已配准的不同模态医学图像的融合.实验结果表明,融合后的图像能保持各输入通道图像中显著的有意义细节和结构信息,有效改善增强图像的视觉效果.更多还原

【Abstract】 Techniques of image enhancement aim at improving the interpretability or perception of information in images for human viewers, eliminating or attenuating unneeded information, or providing better input for other automated image processing techniques. The design of en-hancement methods closely relates to the aim of applications.In recent years, consumer and professional digital cameras are becoming more and more popular, hence enormous amount of image data are being generated. Visual effects of images taken in high dynamic range scenes, dim scenes, or under special lighting conditions are usually not satisfying. Hence, in order to fulfill the requirement for high quality display and print, post enhancement steps are necessary to expand or compress the luminance ratios, or to achieve color constancy. The human visual system excels in perception of scenes in different conditions. Image enhancement methods have been developed to describe the characteristics of human perception, in order to produce enhancement results similar to the human perceived scenes. Several image enhancement approaches are studied in the dissertation, and the advantages and disadvantages are analyzed. Some improvements to these approaches are also proposed. The work of the dissertation includes the following parts:(1) A fast Retinex image enhancement method that models both global and local adaptation of the human visual system is proposed to overcome the deficiency of halo artifacts which tradi-tional center/surround-based Retinex enhancement methods often suffer. Mean Shift filtering is performed on the original image in the illumination estimation step. Due to the capability of discontinuity preservation of adaptive filtering, halo artifacts can be effectively reduced by the proposed method. Experimental results demonstrate that the method can efficiently render high dynamic range images, and the results are compared with previous methods. An acceler-ated implementation of Mean Shift filtering is employed, thus the method is much faster than previous methods.(2) An improved version of Automatic Color Equalization (ACE) is proposed for the en-hancement of image contrast. A local filtering is performed first by taking account the spatial distribution of grayscale or color in the image, and an improved relative lightness appearance function is employed. Then the dynamic range of the image is adjusted for the final result. By keeping the proportion among the RGB channels, the color of the original image is pre-served in processing of color images. Experimental results demonstrate that the contrast can be effectively enhanced and no significant noise is introduced by this method.(3) Recently variational and partial differential equation-based methods have been widely used in the image processing community. An improved image enhancement method in the per-ceptually inspired variational framework is proposed. A spatial adaptive parameter determined by local image features is developed to regulate the contrast enhancement. The minimum of the functional is computed using a gradient descent approach. Experimental results demon-strate that the method can enhancement the contrast of color images and details in dark and flat regions can be effectively improved.(4) Common contrast enhancement methods compress the dynamic range of images while increasing contrast and enhancing the visibility of details in the darker regions. At the same time, details in the brighter regions are usually not enhanced, or even attenuated in some cases. The novelty of the proposed method is to take the human visual perceptual sensitivity into con-sideration. Relying on the observation that the perceived contrast is less at regions with higher local average luminance levels, the idea of relative gradients is introduced and a gradient do-main method that tends to enhance contrast in brighter regions more is proposed. Experimental results demonstrate that details in both brighter and darker regions of the original images are enhanced or preserved.(5) A fusion approach in the gradient domain to combine complementary advantages be-tween results by different image enhancement methods for visualization improvement is pro-posed. A weighted structure tensor is employed to capture significant details of each input channel, and local contrast is incorporated in the design of fusion weights. The target gradi-ent field is obtained from the structure tensor in the Frobenius norm sense. An energy term related to perceptual enhancement is incorporated into the fusion model. Experimental results demonstrate that the fused image can preserve significant detail and structural information of each input image, and the visual effect is improved. Several other applications of the proposed fusion method are presented as well.更多还原