一，图像文件的读取
初学数字图像处理时，图像文件的读取往往是一件麻烦的事情，我们要面对各种各样的图像文件格式，如果仅用C++的fstream库那就必须了解各种图像编码格式，这对于初学图像处理是不太现实的，需要一个能帮助轻松读取各类图像文件的库。在Win32平台上GDI+(C++)是不错的选择，不光使用上相对于Win32 GDI要容易得多，而且也容易移植到.Net平台上的GDI+。
Gdiplus::Bitmap类为我们提供了读取各类图像文件的接口，Bitmap::LockBits方法产生的BitmapData类也为我们提供了高速访问图像文件流的途径。这样我们就可以将精力集中于图像处理算法的实现，而不用关心各种图像编码。具体使用方式请参考MSDN中GDI+文档中关于Bitmap类和BitmapData类的说明。另外GDI+仅在Windows XP/2003上获得直接支持，对于Windows 2000必须安装相关DLL，或者安装有Office 2003，Visual Studio 2003 .Net等软件。
二，空间域图像处理算法框架
(1) 在空间域图像处理中，对于一个图像我们往往需要对其逐个像素的进行处理，对每个像素的处理使用相同的算法（或者是图像中的某个矩形部分）。即，对于图像f(x,y),其中0≤x≤M,0≤y≤N,图像为M*N大小，使用算法algo，则f(x,y) = algo(f(x,y))。事先实现一个算法框架，然后再以函数指针或函数对象(functor,即实现operator()的对象)传入算法，可以减轻编程的工作量。
如下代码便是一例:
#ifndef PROCESSALGO_H
#define PROCESSALGO_H
#include <windows.h>
#include <Gdiplus.h>
namespace nsimgtk
{
template <typename pixelType, Gdiplus::PixelFormat pixelFormat, class Processor>
bool ProcessPixelsOneByOne(Gdiplus::Bitmap* const p_bitmap, Processor processor, unsigned int x, unsigned int y,
unsigned int width, unsigned int height)
{
if (p_bitmap == NULL)
{
return false;
}
if ((width + x > p_bitmap->GetWidth()) || (height + y >p_bitmap->GetHeight()))
{
return false;
}
Gdiplus::BitmapData bitmapData;
Gdiplus::Rect rect(x, y, width,height);
if (p_bitmap->LockBits(&rect, Gdiplus::ImageLockModeWrite, pixelFormat, &bitmapData) != Gdiplus::Ok)
{
return false;
}
pixelType *pixels = (pixelType*)bitmapData.Scan0;
for (unsigned int row=0; row<height; ++row)
{
for (unsigned int col=0; col<width; ++col)
{
processor(&pixels[col+row*bitmapData.Stride/sizeof(pixelType)]);
}
}
if (p_bitmap->UnlockBits(&bitmapData) != Gdiplus::Ok)
{
return false;
}
return true;
}
}
#endif
ProcessPixelsOneByOne函数可以对图像中从(x,y)位置起始，width*height大小的区域进行处理。模板参数pixelType用于指定像素大小，例如在Win32平台上传入unsigned char即为8位，用于8阶灰度图。模板参数Processor为图像处理算法实现，可以定义类实现void operator(pixelType *)函数，或者传入同样接口的函数指针。
如下便是一些算法示例(说明见具体注释):
#ifndef SPATIALDOMAIN_H
#define SPATIALDOMAIN_H
#include <cmath>
#include <string>
namespace nsimgtk
{
// 8阶灰度图的灰度反转算法
class NegativeGray8
{
public:
void operator()(unsigned char *const p_value)
{
*p_value ^= 0xff;
}
};
// 8阶灰度图的Gamma校正算法
class GammaCorrectGray8
{
private:
unsigned char d_s[256];
public:
GammaCorrectGray8::GammaCorrectGray8(double c, double gamma);
void operator()(unsigned char*const p_value)
{
*p_value = d_s[*p_value];
}
};
// 8阶灰度图的饱和度拉伸算法
class ContrastStretchingGray8
{
private:
unsigned char d_s[256];
public:
ContrastStretchingGray8::ContrastStretchingGray8(double a1, double b1, unsigned int x1,
double a2, double b2, unsigned int x2, double a3, double b3);
void operator()(unsigned char*const p_value)
{
*p_value = d_s[*p_value];
}
};
// 8阶灰度图的位平面分割，构造函数指定位平面号
class BitPlaneSliceGray8
{
private:
unsigned char d_s[256];
public:
BitPlaneSliceGray8(unsigned char bitPlaneNum);
void operator()(unsigned char* const p_value)
{
*p_value = d_s[*p_value];
}
};
}
#endif
// 上述类中各构造函数的实现代码，应该分在另一个文件中，此处为说明方便，一并列出
#include "SpatialDomain/spatialDomain.h"
namespace nsimgtk
{
GammaCorrectGray8::GammaCorrectGray8(double c, double gamma)
{
double temp;
for (unsigned int i=0; i<256; ++i)
{
temp = ceil(c * 255.0 * pow(double(i)/255.0, gamma));
d_s[i] = unsigned char(temp);
}
}
ContrastStretchingGray8::ContrastStretchingGray8(double a1, double b1, unsigned int x1,
double a2, double b2, unsigned int x2, double a3, double b3)
{
if (x1 > 255 || x2 > 255 || x1 > x1)
{
for (unsigned int i=0; i<256; ++i)
d_s[i] = i;
}
else
{
double tmp;
for (unsigned int i=0; i<x1; ++i)
{
tmp = ceil(a1*double(i)+b1);
d_s[i] = (unsigned char)tmp;
}
for (unsigned int i=x1; i<x2; ++i)
{
tmp = ceil(a2*double(i)+b2);
d_s[i] = (unsigned char)tmp;
}
for (unsigned int i=x2; i<256; ++i)
{
tmp = ceil(a3*double(i)+b3);
d_s[i] = (unsigned char)tmp;
}
}
}
BitPlaneSliceGray8::BitPlaneSliceGray8(unsigned char bitPlaneNum)
{
unsigned char bitMaskArray[8] =
{
0x01, 0x02, 0x04, 0x08,
0x10, 0x20, 0x40, 0x80
};
for (unsigned int i=0; i<256; ++i)
{
unsigned char tmp = i;
tmp &= bitMaskArray[bitPlaneNum];
tmp = (tmp >> bitPlaneNum) * 255;
d_s[i] = tmp;
}
}
}
(2) 直方图在GDI+1.0中没有获得支持，我们必须自行实现。直方图相关的处理在数字图像处理中占有重要地位，可以通过它获取图像灰度级的统计信息，且可以通过直方图进行一些重要的图像增强技术，如直方图均衡化，直方图规定化，基本全局门限等。
下面是获取8阶图像直方图的算法实现：
namespace nsimgtk
{
bool GetHistogramNormalizeGray8(Gdiplus::Bitmap * const p_bitmap, float *histogramArray)
{
if (p_bitmap == NULL || histogramArray == NULL)
{
return false;
}
Gdiplus::BitmapData bitmapData;
Gdiplus::Rect rect(0, 0, p_bitmap->GetWidth(), p_bitmap->GetHeight());
if (p_bitmap->LockBits(&rect, Gdiplus::ImageLockModeRead, PixelFormat8bppIndexed, &bitmapData) != Gdiplus::Ok)
{
return false;
}
unsigned char *pixels = (unsigned char*)bitmapData.Scan0;
unsigned int histogram[256];
for (int i=0; i<256; ++i)
{
histogram[i] = 0;
}
for (unsigned int row=0; row<p_bitmap->GetWidth(); ++row)
{
for (unsigned int col=0; col<p_bitmap->GetHeight(); ++col)
{
++histogram[pixels[col+row*bitmapData.Stride]];
}
}
const unsigned int totalPixels = p_bitmap->GetWidth() * p_bitmap->GetHeight();
for (int i=0; i<256; ++i)
{
histogramArray[i] = float(histogram[i]) / float(totalPixels);
}
if (p_bitmap->UnlockBits(&bitmapData) != Gdiplus::Ok)
{
return false;
}
return true;
}
}
在获取直方图后(即上面算法的第二个参数)，再将其作为参数传入下面的对象的构造函数，然后以该对象为仿函数传入ProcessPixelsOneByOne即可实现8阶图像直方图均衡化:
#ifndef SPATIALDOMAIN_H
#define SPATIALDOMAIN_H
#include <cmath>
#include <string>
namespace nsimgtk
{
// 8阶灰度图的直方图均衡化
class HistogramEqualizationGray8
{
private:
unsigned char d_s[256];
public:
HistogramEqualizationGray8(const float *const histogramArray);
void operator()(unsigned char *const p_value)
{
*p_value = d_s[*p_value];
}
};
}
#endif
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
#include "SpatialDomain/spatialDomain.h"
namespace nsimgtk
{
HistogramEqualizationGray8::HistogramEqualizationGray8(const float *const histogramArray)
{
if (histogramArray != NULL)
{
float sum = 0.0;
for (int i=0; i<256; ++i)
{
sum += histogramArray[i];
d_s[i] = unsigned char(sum * 255);
}
}
}
}
(3)空间域滤波器，滤波器是一个m*n大小的掩模，其中m,n均为大于1的奇数。滤波器逐像素地通过图像的全部或部分矩形区域，然后逐像素地对掩模覆盖下的像素使用滤波器算法获得响应，将响应赋值于当前像素即掩模中心像素，另外滤波器算法使用中将会涉及到图像边缘的问题，这可以对边缘部分掩模使用补零法补齐掩模下无像素值的区域，或者掩模的移动范围以不越出图像边缘的方式移动，当然这些处理方法都会给图像边缘部分带来不良效果，但是一般情况下，图像边缘部分往往不是我们关注的部分或者没有重要的信息。
下面的滤波器算法框架SpatialFilterAlgo即以补零法(zero-padding)实现:
#ifndef SPATIALFILTER_H
#define SPATIALFILTER_H
#include <vector>
#include <numeric>
#include <algorithm>
#include <gdiplus.h>
#include <fstream>
#include <cmath>
namespace nsimgtk
{
template <typename pixelType, Gdiplus::PixelFormat pixelFormat, class FilterMask>
bool SpatialFilterAlgo(Gdiplus::Bitmap* const p_bitmap, FilterMask filterMask, unsigned int x, unsigned int y,
unsigned int width, unsigned int height)
{
if (p_bitmap == NULL)
{
return false;
}
if ((width + x > p_bitmap->GetWidth()) || (height + y >p_bitmap->GetHeight()))
{
return false;
}
Gdiplus::BitmapData bitmapData;
Gdiplus::Rect rect(x, y, width,height);
if (p_bitmap->LockBits(&rect, Gdiplus::ImageLockModeWrite, pixelFormat, &bitmapData) != Gdiplus::Ok)
{
return false;
}
pixelType *pixels = (pixelType*)bitmapData.Scan0;
const unsigned int m = filterMask.d_m; // mask's width
const unsigned int n = filterMask.d_n; // mask's height
std::vector<pixelType> tmpImage((m-1+width)*(n-1+height)); // extend image to use zero-padding
// copy original bitmap to extended image with zero-padding method
for (unsigned int row=0; row<height; ++row)
{
for (unsigned int col=0; col<width; ++col)
{
tmpImage[(col+m/2)+(row+n/2)*(bitmapData.Stride/sizeof(pixelType)+m-1)] =
pixels[col+row*bitmapData.Stride/sizeof(pixelType)];
}
}
// process every pixel with filterMask
for (unsigned int row=0; row<height; ++row)
{
for (unsigned int col=0; col<width; ++col)
{
// fill the "m*n" mask with the current pixel's neighborhood
for (unsigned int i=0; i<n; ++i)
{
for (unsigned int j=0; j<m; ++j)
{
filterMask.d_mask[i*m+j] = tmpImage[(col+j)+(row+i)*(bitmapData.Stride/sizeof(pixelType)+m-1)];
}
}
// replace the current pixel with filter mask's response
pixels[col+row*bitmapData.Stride/sizeof(pixelType)] = filterMask.response();
}
}
if (p_bitmap->UnlockBits(&bitmapData) != Gdiplus::Ok)
{
return false;
}
return true;
}
}
#endif
其中模板参数FilterMask即为滤波掩模算法。通常的滤波算法有均值滤波器，可以模糊化图像，去除图形中的细节部分，使得我们可以关注图像中较为明显的部分，均值滤波器用于周期性噪声。中值滤波器用于图像中存在椒盐噪声也即脉冲噪声的情况下。另外有基于一阶微分的Sobel梯度算子和基于两阶微分的拉普拉斯算子，它们往往被用于边缘检测中。
下面是一些滤波器算法的具体实现，所以滤波器算法都应该实现pixelType response()函数以及有d_mask,d_m,d_n成员，这可以通过继承__filteMask类获得(不需要付出虚函数代价)。
#ifndef SPATIALFILTER_H
#define SPATIALFILTER_H
#include <vector>
#include <numeric>
#include <algorithm>
#include <gdiplus.h>
#include <fstream>
#include <cmath>
namespace nsimgtk
{
// 滤波器掩模的基类，提供掩模大小d_m,d_n，掩模覆盖下的m*n个像素值d_mask
// others filterMask should inherit it
template <typename pixelType>
struct __filterMask
{
const unsigned int d_m;
const unsigned int d_n;
// image's pixels under the m*n filter mask
std::vector<pixelType> d_mask;
// filter mask's width and heigh must be a odd, if not, it will plus one for the width or the height
__filterMask(unsigned int m, unsigned int n)
: d_m(m%2 ? m:m+1), d_n(n%2 ? n:n+1), d_mask(d_m*d_n)
{
}
};
// 掩模权值为全1的均值滤波器
template <typename pixelType>
class averagingFilterMaskSp
: public __filterMask<pixelType>
{
public:
averagingFilterMaskSp(unsigned int m, unsigned int n)
: __filterMask<pixelType>(m, n)
{ }
pixelType response()
{
return std::accumulate(d_mask.begin(), d_mask.end(), 0) / (d_m * d_n);
}
};
// 可自定义掩模权值的均值滤波器
template <typename pixelType>
class averagingFilterMask
: public __filterMask<pixelType>
{
private:
std::vector<pixelType> d_weight; // weights' vector(m*n)
int d_weight_sum; // all weights' sum
public:
averagingFilterMask(unsigned int m, unsigned int n, const std::vector<pixelType>& weightVec)
: __filterMask<pixelType>(m, n), d_weight(weightVec)
{
if (weightVec.size() != d_mask.size())
{
// if weight's size isn't equal to mask's size, it will change filter mask as a special filter mask
d_weight.resize(d_mask.size(), 1);
}
d_weight_sum = std::accumulate(d_weight.begin(), d_weight.end(), 0);
}
pixelType response()
{
return std::inner_product(d_mask.begin(), d_mask.end(), d_weight.begin(), 0) / d_weight_sum;
}
};
// 中值滤波器
template <typename pixelType>
class medianFilterMask
: public __filterMask<pixelType>
{
public:
medianFilterMask(unsigned int m, unsigned int n)
: __filterMask<pixelType>(m, n)
{ }
pixelType response()
{
std::sort(d_mask.begin(), d_mask.end());
return d_mask[d_mask.size()/2];
}
};
// 3*3拉普拉斯滤波器
// the mask is: [0 1 0 [0 -1 0
// 1 -5 1 or -1 5 -1
// 0 1 0] 0 -1 0]
// if pixel's brightness is less than min, set it to min
// if pixel's brightness is larger than max, set it to max
template <typename pixelType, pixelType min, pixelType max>
class laplacianFilter
: public __filterMask<pixelType>
{
public:
laplacianFilter()
: __filterMask<pixelType>(3, 3)
{ }
pixelType response()
{
int ret = (int)(5*(int)d_mask[4]) - ((int)d_mask[5]+d_mask[3]+d_mask[1]+d_mask[7]);
if (ret < min)
ret = min;
if (ret > max)
ret = max;
return ret;
}
};
// 3*3Sobel滤波器
// the mask is: [-1 -2 -1 [-1 0 1
// 0 0 0 and -2 0 2
// 1 2 1] -1 0 1]
// if pixel's brightness is larger than max, set it to max
template <typename pixelType, pixelType max>
class sobelFilter
: public __filterMask<pixelType>
{
public:
sobelFilter()
: __filterMask<pixelType>(3, 3)
{ }
pixelType response()
{
int ret = ::abs(d_mask[6]+2*d_mask[7]+d_mask[8]-d_mask[0]-2*d_mask[1]-d_mask[2])
+ ::abs(d_mask[2]+2*d_mask[5]+d_mask[8]-d_mask[0]-2*d_mask[3]-d_mask[6]);
if (ret > max)
ret = max;
return ret;
}
};
}
#endif
