面向人眼探测识别的图像优化方法研究

CORC > 沈阳自动化研究所 > 中国科学院沈阳自动化研究所 > 光电信息技术研究室

题名	面向人眼探测识别的图像优化方法研究
作者	蔡铁峰
学位类别	博士
答辩日期	2015-05-28
授予单位	中国科学院沈阳自动化研究所
授予地点	中国科学院沈阳自动化研究所
导师	朱枫
关键词	图像增强人眼视觉特性伪彩色探测识别图像优化
其他题名	Study on Image Optimization methods for Objects Detection and Recognition by Human Eyes
学位专业	模式识别与智能系统
中文摘要	图像包含了景物的很多原始信息，但人眼生理特性决定了人眼对亮度或颜色的分辨能力远不能达到常见数字图像本身具有的分辨力，使得图像中的景物信息可能无法被人眼完全感知到。图像增强通过改变图像灰度值或颜色值,使更多景物信息被人眼感知到,图像优化是图像增强方法中的一类，它的特点是有明确的优化目标，并寻找或求解出那幅优化目标值最佳的图像。现有的图像增强方法有很多,但极少具有明确的优化目标，也没有使得到的图像在一定意义上确保面向人眼探测识别最佳。本文面向人眼探测识别，开展图像优化方法研究，具有重要的理论意义与实际应用价值。图像优化需要改变像素灰度值或颜色值，就有可能改变图像中面向人眼探测识别的信息，为了保真图像面向人眼探测识别的信息，就要在图像优化过程中遵循一定的保真约束。其中，图像信息保真与图像可以优化的空间是处于博弈的两方，图像信息保真约束要求严格就会使图像优化空间小，而要争取到更大的图像优化空间就需要放松图像信息保真约束。成像系统能获得的图像一般可以分为灰度图像与彩色图像，而灰度图像通过灰度颜色映射生成伪彩色图像可以使更多灰度图像信息被人眼感知到。本论文面向人眼探测识别，分别在灰度图像在灰度空间内的优化、灰度图像的伪彩色优化和彩色图像优化这三个方面开展了研究。本论文主要研究内容如下:（1）灰度图像中，相邻像素间是否存在灰度差异往往与景物的轮廓或纹理直接相关，而景物的轮廓与纹理信息是人眼完成探测识别的特别重要依据。然而，当相邻像素间存在灰度差异但灰度差小时，人眼感知不到此灰度差异，也就感知不到它所包含的图像信息。本文提出了一种严格信息保真约束下的灰度图像优化方法，该方法要求全图中任意像素之间灰度值不倒序，在此约束下以图像中最多相邻像素间灰度差异被人眼感知到为优化目标，依托于动态规划算法得到优化目标值最大的那幅图像。为了使灰度图像在灰度空间有大的优化空间，放弃全图任意像素之间灰度值不倒序约束，仅要求邻近像素灰度值人眼感知不倒序，在此约束下提出了一种灰度图像优化方法，以图像中所有存在灰度差异的相邻像素对的灰度差异人眼可感知度最大为优化目标，通过精细的多轮调整像素灰度值，求解出优化目标值最大图像。（2）人眼能分辨的颜色数量远多于人眼能分辨的灰度级数,从而灰度图像的伪彩色优化可以使更多面向人眼探测识别的图像信息被人眼感知到。本文提出了一种严格信息保真约束下的灰度图像伪彩色方法，该方法要求全图中所有像素间灰度关系与对应伪彩色图像中颜色属性关系满足相同的映射准则，以人可以通过伪彩色图像最好地获得原灰度图像包含的图像信息为优化目标，在确保不同灰度级映射到人眼可分辨的不同颜色的前提下通过迭代寻优求得优化目标值最佳伪彩色图像。为了使灰度图像的伪彩色优化有更大的空间，放松图像信息保真约束，仅要求邻近像素灰度关系与对应伪彩色图像中颜色属性关系满足相同映射准则，以人可以通过伪彩色图像最好地获得原灰度图像包含的图像信息为优化目标，通过多轮依次调整像素亮度值、饱和度值与色调值得到优化目标值最佳伪彩色图像。（3）在彩色图像中，相邻像素间是否存在颜色差异往往直接与景物的轮廓与纹理相关，包含重要的面向人眼探测识别的图像信息。当相邻像素存在颜色差异但颜色差小时，人眼感知不到此颜色差异。本文提出了一种严格信息保真约束下的彩色图像优化方法，该方法要求全图中任意像素色调值、饱和度值不变且任意像素之间亮度值保序，在此约束下以图像中所有相邻像素间被人眼感知到的颜色差异最大为优化目标，并通过寻优得到优化目标值最大的那幅图像。为了使彩色图像有大的优化空间，放弃全图任意像素之间灰度值保序的约束，仅在邻近像素灰度值人眼感知不倒序约束下，提出了一种彩色图像优化方法，以图像中所有存在颜色差异的相邻像素对的人眼感知颜色差异最大为优化目标，通过精细的多轮调整像素亮度值，求解出优化目标值最大图像。
索取号	TP391.41/C14/2015
英文摘要	The original image contains a lot of information of the scene, but the physiological characteristics of the human eye determines the color or brightness resolving power of the human eyes cannot meet the common digital image itself with the resolving power, and the scene information of the image may not be fully perceived by the human eyes. Image enhancement can make more the scenery information perceived by changing the image gray value or color value. Image optimization is a kind of image enhancement method, which is characterized by the optimization of clear targets and getting the image with optimal image under the optimization targets. There are a lot of existing image enhancement methods, but rarely with explicit optimization targets, also did not get the image in a certain sense to ensure ensure optimization for objects detection and recognition by human eyes. Image optimization for objects detection and recognition by human eyes will be studied, which is with important theoretical significance and practical application value.Image optimization needs to change the gray value or color value of the pixels, and it is possible to change the image information for objects detection and recognition(IIODR). For the fidelity of IIODR, image optimization should follow certain IIODR fidelity constraints in the optimization process. The image information fidelity and the image optimization performance are two parties in a game. the strict IIODR fidelity will make the image optimization space be small, and greater space for image optimization needs to relax the image information fidelity constraint. The images obtained through imaging system can generally be divided into gray images and color images, and the pseudo color images from gray images by gray-color transforming can make more IIODR of the gray image perceived by the human eyes. The three aspects that gray image optimization in gray space, gray image optimization by pseudo-color coding and color image optimization are studied. The main contents of this dissertation are as follows:(1) For the gray image, gray difference between adjacent pixels are often directly related with objects contour or texture , and contour and texture information of the objects is particularly important for the objects detection and recognition by the human eyes. However, when there is gray value difference between neighboring pixels but the difference is small, the difference will not be perceived, also the contained image information will not be perceived. This dissertation proposed a gray image optimization method with strict information fidelity constraints, and this method requires the gray values of any pixel pair in the image to be unreversed. Under this constraint, the optimization target is set to be maximizing the amount of gray value difference perceivable pixel pairs. The optimal image is found though dynamic programming. In order to increase the image optimization space, give up the gray value unreverse constraint between all pixel pairs, but only gray values among adjacent pixels are required to be not reversed. Under this constraint, an optimization method is proposed, the method will maximize the perception degrees of gray value difference among adjacent pixels through multiple rounds of fine adjustment of pixel gray value.(2) The number of distinguishable colors by human eyes is far more than the number of distinguishable gray levels, and gray image pseudo color optimization can make more IIODR perceived. This dissertation presents a gray image pseudo-color optimization method under strict fidelity of IIODR. This method requires the gray-to-pseudo-color mapping relationship should obey the same rule, and to make the most IIODR of the gray image obtained from the pseudo-color image with ensuring different gray levels should map to distinguishable colors. In order to increase the pseudo-color optimization space, relax the image information fidelity constraint, and only the gray-to-pseudo-color mapping relationship among adjacent pixels should obey the same rule. The proposed method can make the most IIODR of the gray image obtained from the pseudo-color image by recursively adjusting the brightness value, hue value, saturation value of the pixels.(3) In the color image, the color difference among adjacent pixels is often directly associated with the texture and contour of the objects, which include IIODR. When there is color difference between a pair of adjacent pixels but the color difference is small, the difference will not be perceived by human eyes. This dissertation proposes a color image optimization method with strict information fidelity constraints, this method keep the hue and saturation values of any pixels unchanged and requires keep the brightness order among any pixels in the image. Under the constraint, the method maximize the perceived color difference among all adjacent pixels in the image. In order to increase the color image optimization space, give up keeping the brightness order among all pixels, but only the brightness order among adjacent pixels. Under the constraint, a color image optimization method is proposed to maximize the perceived color difference among adjacent pixels by recursively adjustment of brightness values of the pixels.
语种	中文
产权排序	1
页码	104页
内容类型	学位论文
源URL	[http://ir.sia.ac.cn/handle/173321/16778]
专题	沈阳自动化研究所_光电信息技术研究室
推荐引用方式 GB/T 7714	蔡铁峰. 面向人眼探测识别的图像优化方法研究[D]. 中国科学院沈阳自动化研究所. 中国科学院沈阳自动化研究所. 2015.