前言
这篇论文实际上也是《快速 ACE 算法 及其在 图 像拼接 中 的 应 用》这篇论文中的快速ACE算法,我用C++实现了,现在放出来。
算法原理
在介绍中,提到了,高动态图像是指在一幅图像中,既有明亮的区域又有阴影区域,为了使细节清晰,需要满足以下几点:
(1)对动态范围具有一定的压缩能力
(2)对亮暗区域的细节有一定的显示能力
(3)满足(1),(2)的条件下不破坏图像的清晰度
Rizzi等根据Retinex理论提出自动颜色均衡算法,该算法考虑了图像中颜色和亮度的空间位置关系,进行局部的自适应滤波,实现具有局部和非线性特征的图像亮度,色彩与对比度调整,同时满足灰度世界理论和白斑点假设。
算法步骤
- 对图像进行色彩/空域调整,完成图像的色差矫正,得到空域重构图像
Rc(p)=∑j∈Subset,j̸=pd(p,j)r(Ic(p)−Ic(j)),(1),其中 Rc是中间结果, Ic(p)−Ic(j)为2个点的亮度差, d(p,j)表示距离度量函数, r(∗)为亮度表现用函数,需要是奇函数,这一步可以适应局部图像对比度, r(∗)可以放大较小的差异,并丰富大的差异,根据局部内容扩展或者压缩动态范围。一般得, r(∗)为: r(x)=⎩⎨⎧1,x<−Tx/T,−T<=x<=T−1,x>T⎭⎬⎫ - 对矫正后的图像进行动态扩展,一种简单的线性扩展为: Oc(p)=round[127.5+scRc(p)],其中 sc是 [(mc,0),(Mc,255)]的斜率,其中 Mc=maxp[Rc(p)],(3), mc=minp[Rc(p)],还可以将其映射到 [0,255]的空间中: L(x)=maxR−minRR(x)−minR
算法改进
我没看懂这部分,暂时就不写了。
C++源码实现
#include <stdio.h>
#include <iostream>
#include <immintrin.h>
#include <opencv2/opencv.hpp>
#include <opencv2/core/core.hpp>
#include <opencv2/ml/ml.hpp>
#include "opencv2/highgui/highgui.hpp"
using namespace cv;
using namespace cv::ml;
using namespace std;
namespace ACE {
//Gray
Mat stretchImage(Mat src) {
int row = src.rows;
int col = src.cols;
Mat dst(row, col, CV_64FC1);
double MaxValue = 0;
double MinValue = 256.0;
for (int i = 0; i < row; i++) {
for (int j = 0; j < col; j++) {
MaxValue = max(MaxValue, src.at<double>(i, j));
MinValue = min(MinValue, src.at<double>(i, j));
}
}
for (int i = 0; i < row; i++) {
for (int j = 0; j < col; j++) {
dst.at<double>(i, j) = (1.0 * src.at<double>(i, j) - MinValue) / (MaxValue - MinValue);
if (dst.at<double>(i, j) > 1.0) {
dst.at<double>(i, j) = 1.0;
}
else if (dst.at<double>(i, j) < 0) {
dst.at<double>(i, j) = 0;
}
}
}
return dst;
}
Mat getPara(int radius) {
int size = radius * 2 + 1;
Mat dst(size, size, CV_64FC1);
for (int i = -radius; i <= radius; i++) {
for (int j = -radius; j <= radius; j++) {
if (i == 0 && j == 0) {
dst.at<double>(i + radius, j + radius) = 0;
}
else {
dst.at<double>(i + radius, j + radius) = 1.0 / sqrt(i * i + j * j);
}
}
}
double sum = 0;
for (int i = 0; i < size; i++) {
for (int j = 0; j < size; j++) {
sum += dst.at<double>(i, j);
}
}
for (int i = 0; i < size; i++) {
for (int j = 0; j < size; j++) {
dst.at<double>(i, j) = dst.at<double>(i, j) / sum;
}
}
return dst;
}
Mat NormalACE(Mat src, int ratio, int radius) {
Mat para = getPara(radius);
int row = src.rows;
int col = src.cols;
int size = 2 * radius + 1;
Mat Z(row + 2 * radius, col + 2 * radius, CV_64FC1);
for (int i = 0; i < Z.rows; i++) {
for (int j = 0; j < Z.cols; j++) {
if((i - radius >= 0) && (i - radius < row) && (j - radius >= 0) && (j - radius < col)) {
Z.at<double>(i, j) = src.at<double>(i - radius, j - radius);
}
else {
Z.at<double>(i, j) = 0;
}
}
}
Mat dst(row, col, CV_64FC1);
for (int i = 0; i < row; i++) {
for (int j = 0; j < col; j++) {
dst.at<double>(i, j) = 0.f;
}
}
for (int i = 0; i < size; i++) {
for (int j = 0; j < size; j++) {
if (para.at<double>(i, j) == 0) continue;
for (int x = 0; x < row; x++) {
for (int y = 0; y < col; y++) {
double sub = src.at<double>(x, y) - Z.at<double>(x + i, y + j);
double tmp = sub * ratio;
if (tmp > 1.0) tmp = 1.0;
if (tmp < -1.0) tmp = -1.0;
dst.at<double>(x, y) += tmp * para.at<double>(i, j);
}
}
}
}
return dst;
}
Mat FastACE(Mat src, int ratio, int radius) {
int row = src.rows;
int col = src.cols;
if (min(row, col) <= 2) {
Mat dst(row, col, CV_64FC1);
for (int i = 0; i < row; i++) {
for (int j = 0; j < col; j++) {
dst.at<double>(i, j) = 0.5;
}
}
return dst;
}
Mat Rs((row + 1) / 2, (col + 1) / 2, CV_64FC1);
resize(src, Rs, Size((col + 1) / 2, (row + 1) / 2));
Mat Rf= FastACE(Rs, ratio, radius);
resize(Rf, Rf, Size(col, row));
resize(Rs, Rs, Size(col, row));
Mat dst(row, col, CV_64FC1);
Mat dst1 = NormalACE(src, ratio, radius);
Mat dst2 = NormalACE(Rs, ratio, radius);
for (int i = 0; i < row; i++) {
for (int j = 0; j < col; j++) {
dst.at<double>(i, j) = Rf.at<double>(i, j) + dst1.at<double>(i, j) - dst2.at<double>(i, j);
}
}
return dst;
}
Mat getACE(Mat src, int ratio, int radius) {
int row = src.rows;
int col = src.cols;
vector <Mat> v;
split(src, v);
v[0].convertTo(v[0], CV_64FC1);
v[1].convertTo(v[1], CV_64FC1);
v[2].convertTo(v[2], CV_64FC1);
Mat src1(row, col, CV_64FC1);
Mat src2(row, col, CV_64FC1);
Mat src3(row, col, CV_64FC1);
for (int i = 0; i < row; i++) {
for (int j = 0; j < col; j++) {
src1.at<double>(i, j) = 1.0 * src.at<Vec3b>(i, j)[0] / 255.0;
src2.at<double>(i, j) = 1.0 * src.at<Vec3b>(i, j)[1] / 255.0;
src3.at<double>(i, j) = 1.0 * src.at<Vec3b>(i, j)[2] / 255.0;
}
}
src1 = stretchImage(FastACE(src1, ratio, radius));
src2 = stretchImage(FastACE(src2, ratio, radius));
src3 = stretchImage(FastACE(src3, ratio, radius));
Mat dst1(row, col, CV_8UC1);
Mat dst2(row, col, CV_8UC1);
Mat dst3(row, col, CV_8UC1);
for (int i = 0; i < row; i++) {
for (int j = 0; j < col; j++) {
dst1.at<uchar>(i, j) = (int)(src1.at<double>(i, j) * 255);
if (dst1.at<uchar>(i, j) > 255) dst1.at<uchar>(i, j) = 255;
else if (dst1.at<uchar>(i, j) < 0) dst1.at<uchar>(i, j) = 0;
dst2.at<uchar>(i, j) = (int)(src2.at<double>(i, j) * 255);
if (dst2.at<uchar>(i, j) > 255) dst2.at<uchar>(i, j) = 255;
else if (dst2.at<uchar>(i, j) < 0) dst2.at<uchar>(i, j) = 0;
dst3.at<uchar>(i, j) = (int)(src3.at<double>(i, j) * 255);
if (dst3.at<uchar>(i, j) > 255) dst3.at<uchar>(i, j) = 255;
else if (dst3.at<uchar>(i, j) < 0) dst3.at<uchar>(i, j) = 0;
}
}
vector <Mat> out;
out.push_back(dst1);
out.push_back(dst2);
out.push_back(dst3);
Mat dst;
merge(out, dst);
return dst;
}
}
using namespace ACE;
int main() {
Mat src = imread("F:\\sky.jpg");
Mat dst = getACE(src, 4, 7);
imshow("origin", src);
imshow("result", dst);
waitKey(0);
}
效果图
实现的效果和论文有所偏差,这只是拿来参考一下,对这个感兴趣可以研究作者给出的C++源码。
参考博客
https://blog.csdn.net/piaoxuezhong/article/details/78357815
https://www.cnblogs.com/whw19818/p/5765995.html