OpenCv,局部自适应图像增强(Local Adaptive Contrast Enhancement)
转载自:https://blog.csdn.net/EbowTang/article/details/42373081
一、理论
图像增强算法的基本原则是“降低低频区域,突出高频区域”,以此强化边缘,达到增强的目的。最简单的例子就是通过原始图像减去高斯模糊处理后的图像,就能够将边缘强化出来。
直方图均衡化也是一种非常常见的增强方法。但是为了避免背景的干扰,更倾向于采用“局部”方法进行处理。我们这里着重研究自适应对比度增强(ACE)的相关内容。
ACE的定义和原理看上去还是比较简单的。这里的和都可以根据图像本身计算出来。而则需要单独计算。
可以为单独的常量,或者通过来代替。这里的D是一个全局的值,比如平均值。
二、实现
涉及到局部的运算,自然而然会想到使用卷积的方法。更好的是Opencv提供了专门的函数用来做这个工作—BLUR
文档中写到:
-
bool ACE_Enhance(cv::Mat& src_img, cv::Mat& dst_img, unsigned int half_winSize, double Max_Q);
-
double GetMeanValue(cv::Mat& src_img);
-
double GetVarianceValue(cv::Mat& src_img, double MeanVlaue);
-
//************************************************************************
-
// 函数名称: ACE_Enhance
-
// 访问权限: public
-
// 创建日期: 2016/11/23
-
// 创 建 人:
-
// 函数说明: 单通道增强
-
// 函数参数: cv::Mat & src_img 输入图像
-
// 函数参数: cv::Mat & dst_img 输出图像
-
// 函数参数: unsigned int half_winSize 增强窗口的半窗大小
-
// 函数参数: double Max_Q 最大Q值
-
// 返 回 值: bool
-
//************************************************************************
-
bool ACE_Enhance(Mat& src_img, Mat& dst_img, unsigned int half_winSize, double Max_Q)
-
{
-
if (!src_img.data)
-
{
-
cout << "没有输入图像" << endl;
-
return false;
-
}
-
int rows(src_img.rows);
-
int cols(src_img.cols);
-
unsigned char* pdata = nullptr;
-
unsigned char* pdata1 = nullptr;
-
cv::Mat tmpDstImg(rows, cols, CV_8UC1, cv::Scalar::all(0));
-
for (int i = half_winSize; i < (rows - half_winSize); i++)
-
{
-
pdata = tmpDstImg.ptr<unsigned char>(i);
-
pdata1 = src_img.ptr<unsigned char>(i);
-
for (int j = half_winSize; j < (cols - half_winSize); j++)
-
{
-
cv::Mat tempArea = src_img(cv::Rect(j - half_winSize, i - half_winSize, half_winSize * 2 + 1, half_winSize * 2 + 1)); //截取窗口图像
-
double MeanVlaue = GetMeanValue(tempArea);//平均值
-
double varian = GetVarianceValue(tempArea, MeanVlaue);//均方差
-
if (varian > 0.1)
-
{
-
double cg = 100.0 / sqrt(varian);
-
cg = cg > Max_Q ? Max_Q : cg;
-
double pixelvalue = cg*((double)pdata1[j] - MeanVlaue);
-
int tempValue = MeanVlaue + pixelvalue;
-
tempValue = tempValue > 255 ? 255 : tempValue;
-
tempValue = tempValue < 0 ? 0 : tempValue;
-
pdata[j] = tempValue;
-
}
-
}
-
}
-
dst_img = tmpDstImg;
-
return true;
-
}
-
//************************************************************************
-
// 函数名称: GetMeanValue
-
// 访问权限: public
-
// 创建日期: 2016/11/18
-
// 创 建 人:
-
// 函数说明: 获取图像的平均灰度值
-
// 函数参数: cv::Mat & src_img 输入图像
-
// 返 回 值: double
-
//************************************************************************
-
double GetMeanValue(Mat& src_img)
-
{
-
if (CV_8UC1 != src_img.type())
-
{
-
return -1.0;
-
}
-
int rows(src_img.rows); //height
-
int cols(src_img.cols); //width
-
unsigned char* data = nullptr;
-
double PixelValueSum(0.0); //总共的像素值
-
for (int i = 0; i < rows; i++)
-
{
-
data = src_img.ptr<unsigned char>(i);
-
for (int j = 0; j < cols; j++)
-
{
-
PixelValueSum += (double)data[j];
-
} //计算图像的总共像素值
-
}
-
double result(PixelValueSum / static_cast<double>(rows*cols)); //计算图像的均值
-
return result;
-
}
-
//
-
// 函数名称: GetVarianceValue
-
// 访问权限: public
-
// 创建日期: 2016/11/18
-
// 创 建 人:
-
// 函数说明: 计算图像的均方差
-
// 函数参数: cv::Mat & src_img 输入图像
-
// 函数参数: double MeanVlaue 图像的均值
-
// 返 回 值: double
-
//
-
double GetVarianceValue(Mat& src_img, double MeanVlaue)
-
{
-
if (CV_8UC1 != src_img.type())
-
{
-
return -1.0;
-
}
-
int rows(src_img.rows); //height
-
int cols(src_img.cols); //width
-
unsigned char* data = nullptr;
-
double PixelValueSum(0.0); //总共的像素值
-
for (int i = 0; i < rows; i++)
-
{
-
data = src_img.ptr<unsigned char>(i);
-
for (int j = 0; j < cols; j++)
-
{
-
PixelValueSum += pow((double)(data[j] - MeanVlaue), 2);
-
} //计算图像方差
-
}
-
double result(PixelValueSum / static_cast<double>(rows*cols)); //计算图像的均方差
-
return result;
-
}
程序保留:
-
#include <stdio.h>
-
#include <iostream>
-
#include "fftw3.h"
-
#include <stdlib.h>
-
#include <string.h>
-
#include <limits.h>
-
#include "string"
-
#include "vector"
-
#include <windows.h>
-
#include <opencv2/legacy/legacy.hpp>
-
#include <opencv2/nonfree/nonfree.hpp>//opencv_nonfree模块:包含一些拥有专利的算法,如SIFT、SURF函数源码。
-
#include "opencv2/core/core.hpp"
-
#include "opencv2/features2d/features2d.hpp"
-
#include "opencv2/highgui/highgui.hpp"
-
#include <opencv2/nonfree/features2d.hpp>
-
//参考1:基于“局部标准差”的图像增强(原理、算法、代码)
-
//http://www.cnblogs.com/jsxyhelu/p/4857721.html
-
//参考2:使用局部标准差实现图像的局部对比度增强算法。
-
//http://www.cnblogs.com/Imageshop/p/3324282.html
-
using namespace cv;
-
using namespace std;
-
//点乘法,elementWiseMultiplication
-
Mat matrixWiseMulti(Mat &m1, Mat &m2){
-
Mat dst = m1.mul(m2);//注意是对应矩阵位置的元素相乘
-
return dst;
-
}
-
//ACE算法原理:
-
//ACE算法原理表达式:f(i,j)=Mx(i,j)+G(i,j)*[x(i,j)-Mx(i,j)]
-
//高频成分:x(i,j)-Mx(i,j),x(i,j)表示当前中心像素,Mx(i,j)表示局部平均值
-
//增益系数:G(i,j),可以直接令其为系数C(一般总是大于1)
-
//
-
//图像局部对比度增强算法
-
//float MaxCG:对高频成分的最大增益值
-
//int n:局部半径
-
//int C;对高频的直接增益系数
-
//Mat src:原图像
-
Mat ACE(Mat &src, int C = 3, int n = 3, float MaxCG = 7.5){
-
int rows = src.rows;
-
int cols = src.cols;
-
Mat meanLocal;
-
Mat varLocal;
-
Mat meanGlobal;
-
Mat varGlobal;
-
blur(src.clone(), meanLocal, Size(n, n));//meanMask为图像局部均值
-
imshow("低通滤波", meanLocal);
-
Mat highFreq = src - meanLocal;//高频成分:x(i,j)-Mx(i,j)
-
imshow("高频成分", highFreq);
-
varLocal = matrixWiseMulti(highFreq, highFreq);
-
blur(varLocal, varLocal, Size(n, n)); //varMask为此时为图像局部方差
-
//换算成局部标准差(开根号)
-
varLocal.convertTo(varLocal, CV_32F);
-
for (int i = 0; i < rows; i++){
-
for (int j = 0; j < cols; j++){
-
varLocal.at<float>(i, j) = (float)sqrt(varLocal.at<float>(i, j));
-
}
-
}
-
meanStdDev(src, meanGlobal, varGlobal); //meanGlobal为全局均值 varGlobal为全局标准差,实际均是一个数
-
Mat gainArr =0.5 * meanGlobal / varLocal;//增益系数矩阵:G(i,j),可以直接令其为系数C(一般总是大于1)
-
/*
-
for (int i = 0; i < rows; i++){
-
for (int j = 0; j < cols; j++)
-
cout<<gainArr.at<float>(i, j)<<" " ;
-
cout << endl;
-
if (i == 1)
-
break;
-
}*/
-
//对增益矩阵进行截止
-
for (int i = 0; i < rows; i++){
-
for (int j = 0; j < cols; j++){
-
if (gainArr.at<float>(i, j) > MaxCG){
-
gainArr.at<float>(i, j) = MaxCG;
-
}
-
}
-
}
-
gainArr.convertTo(gainArr, CV_8U);
-
gainArr = matrixWiseMulti(gainArr, highFreq);
-
Mat dst1 = meanLocal + gainArr;
-
imshow("Lee改进的D方法", dst1);
-
Mat dst2 = meanLocal + C*highFreq;//直接利用系数C进行高频成分放大
-
imshow("直接系数C方法", dst2);
-
return dst2;
-
}
-
Mat myACE(Mat &src, int n = 7, float MaxCG = 7.5){
-
int rows = src.rows;
-
int cols = src.cols;
-
Mat dst(src.rows, src.cols, CV_8UC1, Scalar::all(0));
-
if (src.type() == CV_8UC1)
-
int aa = src.type();
-
Mat meanLocal;
-
Mat varLocal;
-
Mat meanGlobal;
-
Mat varGlobal;
-
blur(src.clone(), meanLocal, Size(n, n));//meanMask为图像局部均值
-
Mat highFreq = src - meanLocal;//高频成分:x(i,j)-Mx(i,j)
-
varLocal = matrixWiseMulti(highFreq, highFreq);
-
blur(varLocal, varLocal, Size(n, n)); //varMask为此时为图像局部方差
-
//换算成局部标准差(开根号)
-
varLocal.convertTo(varLocal, CV_32F);
-
for (int i = 0; i < rows; i++){
-
for (int j = 0; j < cols; j++){
-
varLocal.at<float>(i, j) = (float)sqrt(varLocal.at<float>(i, j));
-
}
-
}
-
meanStdDev(src, meanGlobal, varGlobal); //meanGlobal为全局均值 varGlobal为全局标准差,实际均是一个数
-
Mat gainArr = 0.5 * meanGlobal / varLocal;//增益系数矩阵:G(i,j),可以直接令其为系数C(一般总是大于1)
-
//对增益矩阵进行截止
-
for (int i = 0; i < rows; i++){
-
for (int j = 0; j < cols; j++){
-
if (gainArr.at<float>(i, j) > MaxCG){
-
gainArr.at<float>(i, j) = MaxCG;
-
}
-
}
-
}
-
gainArr.convertTo(gainArr, CV_8U);
-
gainArr = matrixWiseMulti(gainArr, highFreq);
-
dst = meanLocal + gainArr;
-
//imshow("Lee改进的D方法", dst);
-
return dst;
-
}
-
int main()
-
{
-
const char* img_path = "oct.bmp";
-
//const char* img_path = "yanjing.jpg";
-
//const char* img_path = "luowen.tif";
-
//const char* img_path = "colors_large.png";
-
//const char* img_path = "input_0.png";
-
//const char* img_path = "a1.jpg";
-
//const char* img_path = "seed.tif";
-
//const char* img_path = "flyman_gray.bmp";
-
//const char* img_path = "CT.bmp";
-
//const char* img_path = "rcc.jpg";
-
Mat src = imread(img_path, 0);
-
imshow("src", src);
-
int n = 50;
-
float MaxCG = 10.5;
-
Mat dst=myACE(src,n,MaxCG);
-
imshow("myACE",dst);
-
waitKey();
-
return 0;
-
}