自适应中值滤波器

传统中值滤波器处理低概率椒盐噪声退化的图像效果令人满意但是对于 P = 0.2 的脉冲噪声退化过的图片的处理效果就很令人失望，会丢失很多细节，由此，自适应中值滤波器诞生，与传统不同的是，自适应中值滤波器尝试保留细节。

冈萨雷斯中给出的过程是：
进程A：

A1 = Z<sub>med</sub>  -Z<sub>min</sub>
A2 = Z<sub>med</sub>  -Z<sub>max</sub>
if（A1 > 0  && A2 < 0 ）
	goto B
else if ( S < S<sub>max</sub>)
	goto A
else
	return Z<sub>med</sub>

进程B：

B1 = Z<sub>xy</sub>  -Z<sub>min</sub>
B2 = Z<sub>xy</sub>  -Z<sub>max</sub>
if（B1 > 0  && B2 < 0 ）
	return Z<sub>xy</sub>
else
	return Z<sub>med</sub>

过程的主要思想就是如果当前像素点灰度值不是脉冲，则保留，当滤波器扩大到指定最大之后则选取中值为结果，可能该处是边缘。

突发奇想既然要尽可能多的保留图像原本细节为什么不把B过程和A过程交换一下逻辑经过试验结果是差不多的0。0

实验图片

在这里插入图片描述

#include<opencv2\opencv.hpp>
#include<iostream>

using namespace std;
using namespace cv;

string st = "E:\\冈萨雷斯OpenCV重写\\DIP3E_CH05_Original_Images\\DIP3E_CH05_Original_Images\\";

//自己尝试
uchar autoMedianFilter(Mat src,int i,int j, int Smax, int Smin,int Snow = -1)
{
	if (Snow == -1){ Snow = Smin; }
	double Zmin;
	double Zmax;
	double Zmed;
	double Zxy;

	double A1;
	double A2;

	double B1;
	double B2;

	Mat dst = src.clone();
	Mat tmp;
	Mat tmp2(Size(pow(Snow,2),1),CV_8UC1);


	tmp = dst(Rect(j-Snow/2, i-Snow/2, Snow, Snow)).clone();
	
	for (int z = 0; z < Snow; z++)
	for (int z1 = 0; z1 < Snow;z1++)
	{
		tmp2.at<uchar>(z*Snow + z1) = tmp.at<uchar>(z, z1);
	}
	cv::sort(tmp2, tmp2, CV_SORT_ASCENDING);

	Zxy = tmp.at<uchar>(Snow / 2, Snow / 2);
	Zmin = tmp2.at<uchar>(0);
	Zmax = tmp2.at<uchar>(Snow*Snow-1);
	Zmed = tmp2.at<uchar>(Snow*Snow / 2 );

	A1 = Zmed - Zmin;
	A2 = Zmed - Zmax;

	B1 = Zxy - Zmin;
	B2 = Zxy - Zmax;
	if (B1 > 0 && B2 < 0)
	{
		return Zxy;
	}
	else
	{
		if (A1>0 && A2 < 0)
		{
			return Zmed;
		}
		else
		{
			Snow += 2;
			if (Snow >= Smax){ return Zmed; }
			else{ return autoMedianFilter(src, i, j, Smax, Smin, Snow); }
		}
		
	}
}
//书中的方法
uchar autoMedianFilter2(Mat src, int i, int j, int Smax, int Smin, int Snow = -1)
{
	if (Snow == -1){ Snow = Smin; }
	double Zmin;
	double Zmax;
	double Zmed;
	double Zxy;

	double A1;
	double A2;

	double B1;
	double B2;

	Mat dst = src.clone();
	Mat tmp;
	Mat tmp2(Size(pow(Snow, 2), 1), CV_8UC1);


	tmp = dst(Rect(j - Snow / 2, i - Snow / 2, Snow, Snow)).clone();

	for (int z = 0; z < Snow; z++)
	for (int z1 = 0; z1 < Snow; z1++)
	{
		tmp2.at<uchar>(z*Snow + z1) = tmp.at<uchar>(z, z1);
	}
	cv::sort(tmp2, tmp2, CV_SORT_ASCENDING);

	Zxy = tmp.at<uchar>(Snow / 2, Snow / 2);
	Zmin = tmp2.at<uchar>(0);
	Zmax = tmp2.at<uchar>(Snow*Snow - 1);
	Zmed = tmp2.at<uchar>(Snow*Snow / 2);

	A1 = Zmed - Zmin;
	A2 = Zmed - Zmax;

	B1 = Zxy - Zmin;
	B2 = Zxy - Zmax;

	if (A1>0 && A2 < 0)
	{
		if (B1 > 0 && B2 < 0)
		{
			return Zxy;
		}
		else
		{
			return Zmed;
		}
	}
	else
	{
		Snow += 2;
		if (Snow >= Smax){ return Zmed;  }
		else{ return autoMedianFilter2(src, i, j, Smax, Smin, Snow); }
	}
}
void main()
{
	Mat srcImage = imread(st + "Fig0514(a)(ckt_saltpep_prob_pt25).tif",0);
	imshow("原图像",srcImage);
	Mat dstImage = srcImage.clone();
	copyMakeBorder(srcImage, dstImage, 7, 7, 7, 7, BORDER_REPLICATE);
	for (int i = 4; i < dstImage.rows-4;i++)
	for (int j = 4; j < dstImage.cols-4; j++)
	{
		//这里要用改变后的图像进行计算 逐步优化，使用初始图片效果不会很理想
		dstImage.at<uchar>(i, j) = autoMedianFilter(dstImage, i, j, 7, 3);
	}
	dstImage = dstImage(Rect(7,7,srcImage.cols,srcImage.rows));
	imshow("自适应中值滤波后1",dstImage);

	copyMakeBorder(srcImage, dstImage, 7, 7, 7, 7, BORDER_REPLICATE);
	for (int i = 4; i < dstImage.rows - 4; i++)
	for (int j = 4; j < dstImage.cols - 4; j++)
	{
		dstImage.at<uchar>(i, j) = autoMedianFilter2(dstImage, i, j, 7, 3);
	}
	dstImage = dstImage(Rect(7, 7, srcImage.cols, srcImage.rows));
	imshow("自适应中值滤波后2", dstImage);
	Mat medianImage = srcImage.clone();
	medianBlur(srcImage, medianImage, 7);
	imshow("中值滤波器7*7",medianImage);
	waitKey(0);
}