cv序列的像素中值::垫

Question

注意：我不是在问中值滤波器。

我有一系列图像让我们说：

std::array<cv::Mat,N> sequence;

我想把所有这些图像融为一体。这一个图像应该满足:

新图像的每个像素都是序列中对应像素的中位数。换句话说：

Result(i,j)=median(sequence[0](i,j), sequence[1](i,j), ..., sequence[N](i,j));

有内置功能吗？最快的方法是什么？

到目前为止，我尝试的是:迭代所有序列中的每个像素，排序，然后取中值，然后将其存储在结果中。然而，它是如此矫枉过正。

Answer 1

如果序列中的图像数为奇数，则可以使用以下技术。

• 从您拥有的图像序列中准备一个N（这是奇数）通道图像
• 将此图像重塑为1通道列向量
• 现在对这个列向量应用一个大小为N的中值过滤器，由于图像是1通道的列向量，过滤器会计算通道的中值（当然会有一些额外的计算对我们没有用）
• 将此过滤后的图像重塑为具有N个通道和原始行数和列数的原始形式
• 选择这个N通道图像的中间通道，其中包含序列的中值图像

下面我用一个简单的代码及其输出说明了上述项目。

个别频道

channel 0:
[1, 1, 1;
  1, 1, 1;
  1, 1, 1]
channel 1:
[2, 2, 2;
  2, 2, 2;
  2, 2, 2]
channel 2:
[3, 3, 3;
  3, 3, 3;
  3, 3, 3]
channel 3:
[4, 4, 4;
  4, 4, 4;
  4, 4, 4]
channel 4:
[5, 5, 5;
  5, 5, 5;
  5, 5, 5]

N=3的输出

3-channel image data:
[1, 2, 3, 1, 2, 3, 1, 2, 3;
  1, 2, 3, 1, 2, 3, 1, 2, 3;
  1, 2, 3, 1, 2, 3, 1, 2, 3]
1-channel column vector image data:
[1; 2; 3; 1; 2; 3; 1; 2; 3; 1; 2; 3; 1; 2; 3; 1; 2; 3; 1; 2; 3; 1; 2; 3; 1; 2; 3]
median of the 1-channel column vector image data:
[1; 2; 2; 2; 2; 2; 2; 2; 2; 2; 2; 2; 2; 2; 2; 2; 2; 2; 2; 2; 2; 2; 2; 2; 2; 2; 3]
reshaped filtered image data:
[1, 2, 2, 2, 2, 2, 2, 2, 2;
  2, 2, 2, 2, 2, 2, 2, 2, 2;
  2, 2, 2, 2, 2, 2, 2, 2, 3]
median image data:
[2, 2, 2;
  2, 2, 2;
  2, 2, 2]

N = 5时的输出

5-channel image data:
[1, 2, 3, 4, 5, 1, 2, 3, 4, 5, 1, 2, 3, 4, 5;
  1, 2, 3, 4, 5, 1, 2, 3, 4, 5, 1, 2, 3, 4, 5;
  1, 2, 3, 4, 5, 1, 2, 3, 4, 5, 1, 2, 3, 4, 5]
1-channel column vector image data:
[1; 2; 3; 4; 5; 1; 2; 3; 4; 5; 1; 2; 3; 4; 5; 1; 2; 3; 4; 5; 1; 2; 3; 4; 5; 1; 2
; 3; 4; 5; 1; 2; 3; 4; 5; 1; 2; 3; 4; 5; 1; 2; 3; 4; 5]
median of the 1-channel column vector image data:
[1; 2; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3
; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 3; 4; 5]
reshaped filtered image data:
[1, 2, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3;
  3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3;
  3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 4, 5]
median image data:
[3, 3, 3;
  3, 3, 3;
  3, 3, 3]

代码:

// number of channels (= number of images in the sequence)
// N MUST BE ODD
const int N = 5;    
// channel data
uchar ch0[] = {1, 1, 1, 1, 1, 1, 1, 1, 1};
uchar ch1[] = {2, 2, 2, 2, 2, 2, 2, 2, 2};
uchar ch2[] = {3, 3, 3, 3, 3, 3, 3, 3, 3};
uchar ch3[] = {4, 4, 4, 4, 4, 4, 4, 4, 4};
uchar ch4[] = {5, 5, 5, 5, 5, 5, 5, 5, 5};
// images
Mat m0 = Mat(3, 3, CV_8U, ch0);
Mat m1 = Mat(3, 3, CV_8U, ch1);
Mat m2 = Mat(3, 3, CV_8U, ch2);
Mat m3 = Mat(3, 3, CV_8U, ch3);
Mat m4 = Mat(3, 3, CV_8U, ch4);
// prepare image sequence
Mat channels[] = {m0, m1, m2, m3, m4};
// put the images into channels of matrix m
Mat m;
merge(channels, N, m);
// reshape data so that we have a single channel column vector as the image
Mat n = m.reshape(1, m.rows * m.cols * m.channels());
// apply median filter to the 1-channel column vector image. filter size must be the number of channels
Mat med;
medianBlur(n, med, N);

cout << N << "-channel image data:" << endl;
cout << m << endl;

cout << "1-channel column vector image data:" << endl;
cout << n << endl;

cout << "median of the 1-channel column vector image data:" << endl;
cout << med << endl;

// reshape the filtered 1-channel column vector image into its original form having N channels
med = med.reshape(N, m.rows);

cout << "reshaped filtered image data:" << endl;
cout << med << endl;

// split the image
split(med, channels);

// extract the middle channel which contains the median image of the sequence
cout << "median image data:" << endl;
cout << channels[(N+1)/2 - 1] << endl;

Answer 2

您可以使用直方图计算每个位置的顺序中值。

假设您使用的是< code>Mat1b图像，每个直方图将有256个值。您需要存储直方图以及所有柱的总和:

struct Hist {
    vector<short> h;
    int count;
    Hist() : h(256, 0), count(0) {};
};

中值是直方图中对应于一半像素count/2的索引。Rosetta Code中的片段：

int i;
int n = hist.count / 2; // 'hist' is the Hist struct at a given pixel location
for (i = 0; i < 256 && ((n -= hist.h[i]) >= 0); ++i);
// 'i' is the median value

添加或删除图像时，您会更新每个像素位置的直方图，并重新计算中值。这个操作相当快，因为不需要排序。

这有一些缺点:

>

这将仅适用于uchar值，否则每个直方图的长度将太大

这种方法将使用大量RAM，因为它需要row x ol直方图。它可能不适用于大图像。

以下实现适用于单通道图像，但可以轻松扩展到3通道。

您可以使用基于两个堆的方法或近似方法。您可以在此处找到详细信息：

• 从整数流中找到运行中位数
• C中的滚动中值算法
• 从增长集合中找到中值

代码:

#include <vector>
#include <opencv2/opencv.hpp>

using namespace std;
using namespace cv;

struct Hist {
    vector<short> h;
    int count;
    Hist() : h(256, 0), count(0) {};
};

void addImage(vector<Mat1b>& images, Mat1b& img, vector<vector<Hist>>& M, Mat1b& med)
{
    assert(img.rows == med.rows);
    assert(img.cols == med.cols);

    for (int r = 0; r < img.rows; ++r) {
        for (int c = 0; c < img.cols; ++c){

            // Add pixel to histogram
            Hist& hist = M[r][c];
            ++hist.h[img(r, c)];
            ++hist.count;

            // Compute median
            int i;
            int n = hist.count / 2;
            for (i = 0; i < 256 && ((n -= hist.h[i]) >= 0); ++i);

            // 'i' is the median value
            med(r,c) = uchar(i);
        }
    }

    // Add image to my list
    images.push_back(img.clone());
}

void remImage(vector<Mat1b>& images, int idx, vector<vector<Hist>>& M, Mat1b& med)
{
    assert(idx >= 0 && idx < images.size());

    Mat1b& img = images[idx];
    for (int r = 0; r < img.rows; ++r) {
        for (int c = 0; c < img.cols; ++c){

            // Remove pixel from histogram
            Hist& hist = M[r][c];
            --hist.h[img(r, c)];
            --hist.count;

            // Compute median
            int i;
            int n = hist.count / 2;
            for (i = 0; i < 256 && ((n -= hist.h[i]) >= 0); ++i);

            // 'i' is the median value
            med(r, c) = uchar(i);
        }
    }

    // Remove image from list
    images.erase(images.begin() + idx);
}

void init(vector<vector<Hist>>& M, Mat1b& med, int rows, int cols)
{
    med = Mat1b(rows, cols, uchar(0));
    M.resize(rows);
    for (int i = 0; i < rows; ++i) {
        M[i].resize(cols);
    }
}


int main()
{
    // Your images... be sure that they have the same size
    Mat1b img0 = imread("path_to_image", IMREAD_GRAYSCALE);
    Mat1b img1 = imread("path_to_image", IMREAD_GRAYSCALE);
    Mat1b img2 = imread("path_to_image", IMREAD_GRAYSCALE);
    resize(img0, img0, Size(800, 600));
    resize(img1, img1, Size(800, 600));
    resize(img2, img2, Size(800, 600));



    int rows = img0.rows;
    int cols = img0.cols;

    vector<Mat1b> images;   // All your images, needed only if you need to remove an image
    vector<vector<Hist>> M; // histograms
    Mat1b med;              // median image

    // Init data strutctures
    init(M, med, rows, cols);

    // Add images. 'med' will be the median image and will be updated each time
    addImage(images, img0, M, med);
    addImage(images, img1, M, med);
    addImage(images, img2, M, med);

    // You can also remove an image from the median computation
    remImage(images, 2, M, med);

    // Hey, same median as img0 and img1 ;D

    return 0;
}