原文地址http://blog.sina.com.cn/s/blog_684c8d630100turx.html
刚开会每周的例会,最讨厌开会了,不过为了能顺利毕业,只能忍了。闲话不多说了,下面把上周学习的一个简单的算法总结一下,以备后面写毕业论文的时候可以参考一下。
一、Census Transform(CT)算法的学习
Census Transform 算法是Ramin Zabih 和 John Woodfill 于1994年在他们的论文《Non-parametric LocalTransforms for Computing VisualCorrespondence》中提出的,正如他们在论文中所说,这是一种非参数变换,主要用来表征图像的局部结构特征,能够比较好的检测到图像中的边缘特征和角点特征,从这篇论文的470次的引用次数来看,CT算法用处还是挺广泛的。下面简要介绍一下CT算法的基本思想:用一个3*3或者5*5的滑动窗口遍历整幅图像,对于每次遍历的位置,以3*3为例,假设某个位置如下图所示
123 | 127 | 129 |
126 | 128 | 129 |
127 | 131 | 130 |
然后比较此窗口中每个像素值(中心除外)与中心像素值的大小,如果比中心像素值小,则比较结果为1,否则为0。由此,得到如下结果:
1 | 1 | 0 |
1 | 0 | |
1 | 0 | 0 |
然后,把此窗口结果组成一个序列:11010100,以此二进制序列表示的值来代替原图像窗口中心点的像素。如此下去,等到窗口滑动完整幅图像,我们就得到原图像做统计变换(CT)之后的图像。
注意:只能作用于灰度图像,对于彩色图像,则需要转换为灰度图之后再操作;为了简单起见,我没有考虑图像的边缘像素值。
下面给出一种用 matlab 实现的版本:
- % *************************************************************************
- % Title: Function-Census Transform of a given Image
- % Author: Siddhant Ahuja
- % Created: May 2008
- % Copyright Siddhant Ahuja, 2008
- % Inputs: Image (var: inputImage), Window size assuming square window (var:
- % windowSize) of 3x3 or 5x5 only.
- % Outputs: Census Tranformed Image (var: censusTransformedImage),
- % Time taken (var: timeTaken)
- % Example Usage of Function: [a,b]=funcCensusOneImage('Img.png', 3)
- % *************************************************************************
- function [censusTransformedImage, timeTaken] = funcCensusOneImage(inputImage, windowSize)
- % Grab the image information (metadata) using the function imfinfo
- try
- imageInfo = imfinfo(inputImage);
- % Since Census Transform is applied on a grayscale image, determine if the
- % input image is already in grayscale or color
- if(getfield(imageInfo,'ColorType')=='truecolor')
- % Read an image using imread function, convert from RGB color space to
- % grayscale using rgb2gray function and assign it to variable inputImage
- inputImage=rgb2gray(imread(inputImage));
- else if(getfield(imageInfo,'ColorType')=='grayscale')
- % If the image is already in grayscale, then just read it.
- inputImage=imread(inputImage);
- else
- error('The Color Type of Input Image is not acceptable. Acceptable color types are truecolor or grayscale.');
- end
- end
- catch
- inputImage = inputImage;
- end
- % Find the size (columns and rows) of the image and assign the rows to
- % variable nr, and columns to variable nc
- [nr,nc] = size(inputImage);
- % Check the size of window to see if it is an odd number.
- if (mod(windowSize,2)==0)
- error('The window size must be an odd number.');
- end
- if (windowSize==3)
- bits=uint8(0);
- % Create an image of size nr and nc, fill it with zeros and assign
- % it to variable censusTransformedImage of type uint8
- censusTransformedImage=uint8(zeros(nr,nc));
- else if (windowSize==5)
- bits=uint32(0);
- % Create an image of size nr and nc, fill it with zeros and assign
- % it to variable censusTransformedImage of type uint32
- censusTransformedImage=uint32(zeros(nr,nc));
- else
- error('The size of the window is not acceptable. Just 3x3 and 5x5 windows are acceptable.');
- end
- end
- % Initialize the timer to calculate the time consumed.
- tic;
- % Find out how many rows and columns are to the left/right/up/down of the
- % central pixel
- C= (windowSize-1)/2;
- for j=C+1:1:nc-C % Go through all the columns in an image (minus C at the borders)
- for i=C+1:1:nr-C % Go through all the rows in an image (minus C at the borders)
- census = 0; % Initialize default census to 0
- for a=-C:1:C % Within the square window, go through all the rows
- for b=-C:1:C % Within the square window, go through all the columns
- if (~(a==0 && b==0)) % Exclude the centre pixel from the calculation,原来是(C+1),现改为0
- census=bitshift(census,1); %Shift the bits to the left by 1
- % If the intensity of the neighboring pixel is less than
- % that of the central pixel, then add one to the bit
- % string
- if (inputImage(i+a,j+b) < inputImage(i,j))
- census=census+1;
- end
- end
- end
- end
- % Assign the census bit string value to the pixel in imgTemp
- censusTransformedImage(i,j) = census;
- end
- end
- % Stop the timer to calculate the time consumed.
- timeTaken=toc;
这是我在网上找到的一个实现的版本,注释比较多,除去注释的话,真正代码没有50行,比较简单,相信大家都可以看的懂。
之前讲了CT算法的实现,下面说一下 MCT 以及 RMCT的实现。
二、Modified Census Transform (MCT)算法
MCT 算法是 CT 算法的一个修改版本,它是由Bernhard Froba 在做人脸检测的时候提出来的,在他2004年发表的论文《Face Detection with theModified Census Transform》中,Bernhard Froba将 CT算法中“滑动窗口中每个像素值与中心位置像素做比较”改为“滑动窗口中每个像素值与整个窗口中像素的均值做比较”,这样,原有的每个3*3的窗口可能产生256种序列(因为没有算中心像素),现在变为可能产生512种序列(其实全0和全1的序列表示的是同样的信息,可以排除一个),也就是做完MCT之后,图像的每个像素值的范围为0——511,这样就能够比较充分的利用3*3的核(至于为什么这么说,可以看看前面提到的那篇论文)。如此来计算的话,则前面例子产生的结果窗口应该为:
1 | 1 | 0 |
1 | 0 | 0 |
1 | 0 | 0 |
对应的二进制序列为:110100100。然后以此作为中心像素点的像素值,循环完毕之后便得到MCT之后的图像。需要注意的一点是,如果要使变换之后的图像得到显示,应该对像素值做一下归一化,使其在0——255之间。
三、Revised Modified Census Transform (RMCT)算法
RMCT 算法其实又是对 MCT的又一次修改,它与 MCT的不同之处仅仅在于一个微小的△m,即:在滑动窗口像素均值上加上一个微小的变量△m=1或者2。其他都是完全一样的。
下面附上这两种修改版统计变换的 C++代码,代码是我自己编的,是基于VS2008和OpenCV2.0的,仅供参考:
- #include "stdafx.h"
- #include "MCT.h"
- #include "highgui.h"
- MCT::MCT()
- {
- window_size = 0;
- }
- MCT::~MCT()
- {
- }
- void MCT::ModifiedCensusTransform(IplImage *input_image, IplImage *mct_image, const int window_size, const int delta )
- {
- CvSize image_size = cvGetSize(input_image);
- int image_width = image_size.width;
- int image_height = image_size.height;
- IplImage *gray_image = cvCreateImage(cvGetSize(input_image), input_image->depth, 1);
- cvSetZero(gray_image);
- if(input_image->nChannels != 1)
- {
- cvCvtColor(input_image, gray_image, CV_RGB2GRAY);
- }
- else
- {
- cvCopy(input_image, gray_image);
- }
- IplImage *modified_image = NULL;
- switch (window_size)
- {
- case 3:
- modified_image = cvCreateImage(image_size, IPL_DEPTH_16U, 1);
- cvSetZero(modified_image);
- break;
- case 5:
- modified_image = cvCreateImage(image_size, IPL_DEPTH_32S, 1);
- cvSetZero(modified_image);
- break;
- default:
- printf("window size must be 3 or 5!\n");
- exit(EXIT_FAILURE);
- }
- CvMat window;
- for(int i = 0; i < image_height - window_size; i++)
- {
- for(int j = 0; j < image_width - window_size; j++)
- {
- unsigned long census = 0;
- CvRect roi = cvRect(j, i, window_size, window_size);
- cvGetSubRect(gray_image, &window, roi);
- CvScalar m = cvAvg(&window, NULL);
- for(int w = 0; w < window_size; w++)
- {
- for(int h = 0; h < window_size; h++)
- {
- census = census << 1; //左移1位
- double tempvalue = cvGetReal2D(&window, w, h);
- if(tempvalue < m.val[0] + delta)
- census += 1;
- }
- }
- cvSetReal2D(modified_image, i, j, census);
- }
- }
- //cvConvertScaleAbs(modified_image, mct_image, 1, 0);
- Normalize(modified_image, mct_image);
- cvReleaseImage(&gray_image);
- cvReleaseImage(&modified_image);
- }
- void MCT::Normalize(IplImage *mct_image, IplImage *nor_image)
- {
- double minv, maxv;
- cvMinMaxLoc(mct_image, &minv, &maxv);
- for (int i = 0; i < mct_image->height; i++)
- {
- for (int j = 0; j < mct_image->width; j++)
- {
- double tempv = cvGetReal2D(mct_image, i, j);
- tempv = (tempv - minv) / (maxv - minv) * 255;
- cvSetReal2D(nor_image, i, j, tempv);
- }
- }
- }
由于算法比较简单,所以没有写注释,应该比较容易理解。
四、CT/MCT/RMCT的应用
目前我读到的几篇论文里面,他们主要用于人脸检测或者面部伪装检测,用于做图像的预处理,这可能是因为CT算法对光照不敏感,可以比较好的排除光照对图像的影响。这里给出用到该算法的几篇paper:
1、《Face Detection with the Modified CensusTransform》(前面提到的那篇)
2、《Adaboost Based Disguised Face Discrimination on EmbeddedDevices》
3、《Disguised-Face Discriminator for Embedded Systems》
OVER!