jashliao 用 VC++ 實現 fanfuhan OpenCV 教學042 ~ opencv-042-二值化圖像OTSU演算法(THRESH_OTSU)
jashliao 用 VC++ 實現 fanfuhan OpenCV 教學042 ~ opencv-042-二值化圖像OTSU演算法(THRESH_OTSU)
資料來源: https://fanfuhan.github.io/
https://fanfuhan.github.io/2019/04/13/opencv-042/
GITHUB:https://github.com/jash-git/fanfuhan_ML_OpenCV
https://github.com/jash-git/jashliao-implements-FANFUHAN-OPENCV-with-VC
★前言:
★主題:
otsu 大津算法介紹:
OTSU算法是由日本學者OTSU於1979年提出的一種對圖像進行二值化的高效算法。
利用閾值將原圖像分成前景,背景兩個圖象。
前景:用n1,csum,m1來表示在當前閾值下的前景的點數,質量矩,平均灰度
背景:用n2, sum-csum,m2來表示在當前閾值下的背景的點數,質量矩,平均灰度
當取最佳閾值時,背景應該與前景差別最大,關鍵在於如何選擇衡量差別的標准,而在otsu算法中這個衡量差別的標准就是最大類間方差,在本程序中類間方差用sb表示,最大類間方差用fmax
otsu 大津算法原理
otsu 大津算法是一種圖像二值化算法,作用是確定將圖像分成黑白兩個部分的閾值。
將圖像背景和前景分成黑白兩類很好理解,但是如何確定背景和前景的二值化界限(閾值)呢?
對於不同的圖像,這個閾值可能不同,這就需要有一種算法來根據圖像的信息自適應地確定這個閾值。
首先,需要將圖像轉換成灰度圖像,255個灰度等級。
可以將圖像理解成255個圖層,每一層分布了不同的像素,這些像素垂直疊加合成了一張完整的灰度圖。
我們的目的就是找到一個合適的灰度值,大於這個值的我們將它稱之為背景(灰度值越大像素越黑),小於這個值的我們將它稱之為前景(灰度值越小像素越白)。
怎么確定這個值就是我們想要的值呢?
這里引入方差的概念,方差越大,相關性越低,黑白越分明。
我們將每一個灰度值之上下之間的像素的方差求出來不就行了嗎?找到方差最大的那個灰度值,那個就是我們想要的二值化分隔閾值。
先定義几個符號代表的意義:
h:圖像的寬度
w:圖像的高度(h*w 得到圖像的像素數量)
t :灰度閾值(我們要求的值,大於這個值的像素我們將它的灰度設置為255,小於的設置為0)
n0:小於閾值的像素,前景
n1:大於等於閾值的像素,背景
n0 + n1 == h * w
w0:前景像素數量占總像素數量的比例
w0 = n0 / (h * w)
w1:背景像素數量占總像素數量的比例
w1 = n1 / (h * w)
w0 + w1 == 1
u0:前景平均灰度
u0 = n0灰度累加和 / n0
u1:背景平均灰度
u1 = n1灰度累加和 / n1
u:平均灰度
u = (n0灰度累加和 + n1灰度累加和) / (h * w) 根據上面的關系
u = w0 * u0 + w1 * u1
g:類間方差(那個灰度的g最大,哪個灰度就是需要的閾值t)
g = w0 * (u0 – u)^2 + w1 * (u1 – u)^2
根據上面的關系,可以推出:(這個一步一步推導就可以得到)
g = w0 * w1 * (u0 – u1) ^ 2
然后,遍曆每一個灰度值,找到這個灰度值對應的 g
找到最大的 g 對應的 t
★C++
// VC_FANFUHAN_OPENCV042.cpp : 定義主控台應用程式的進入點。 // /* // Debug | x32 通用屬性 | C/C++ | | 一般 | | 其他 Include 目錄 -> C:\opencv\build\include | | 連結器 | |一一般 | | 其他程式庫目錄 -> C:\opencv\build\x64\vc15\lib | | |一輸入 | | 其他相依性 -> opencv_world411d.lib;%(AdditionalDependencies) // Releas | x64 組態屬性 | C/C++ | | 一般 | | 其他 Include 目錄 -> C:\opencv\build\include | | 連結器 | |一般 | | 其他程式庫目錄 -> C:\opencv\build\x64\vc15\lib | | |一輸入 | | 其他相依性 -> opencv_world411.lib;%(AdditionalDependencies) */ #include "stdafx.h" #include <iostream> #include <opencv2/opencv.hpp> #include <opencv2/core/core.hpp> #include <opencv2/highgui/highgui.hpp> using namespace std; using namespace cv; void blur_demo(Mat &image, Mat &sum); void edge_demo(Mat &image, Mat &sum); int getblockSum(Mat &sum, int x1, int y1, int x2, int y2, int i); void showHistogram(InputArray src, cv::String StrTitle); void backProjection_demo(Mat &mat, Mat &model); void blur3x3(Mat &src, Mat *det); void add_salt_pepper_noise(Mat &image); void add_gaussian_noise(Mat &image); void USMImage(Mat src, Mat &usm, float fltPar); void pyramid_up(Mat &image, vector<Mat> &pyramid_images, int level); void pyramid_down(vector<Mat> &pyramid_images); void laplaian_demo(vector<Mat> &pyramid_images, Mat &image); void pause() { printf("Press Enter key to continue..."); fgetc(stdin); } int main() { Mat src = imread("../../images/master.jpg");//Mat src = imread("../../images/test.png"); if (src.empty()) { cout << "could not load image.." << endl; pause(); return -1; } else { imshow("input_src", src); showHistogram(src, "Histogram_input_src"); // 转为灰度图像 Mat gray, binary; cvtColor(src, gray, COLOR_BGR2GRAY); //imshow("input_gray", gray); //showHistogram(gray, "Histogram_gray"); Scalar m = mean(gray); int T = mean(src)[0]; for (int i = 0; i < 5; ++i) { // THRESH_BINARY = 0 // THRESH_BINARY_INV = 1 // THRESH_TRUNC = 2 // THRESH_TOZERO = 3 // THRESH_TOZERO_INV = 4 threshold(gray, binary, 0, 255, i| THRESH_OTSU); imshow(format("binary_%d | THRESH_OTSU", i), binary); } waitKey(0); return 0; } return 0; } void laplaian_demo(vector<Mat> &pyramid_images, Mat &image)//拉普拉斯金字塔 { for (int i = pyramid_images.size() - 1; i > -1; --i) { Mat dst; if (i - 1 < 0) { pyrUp(pyramid_images[i], dst, image.size()); subtract(image, dst, dst);//圖像相減 dst = dst + Scalar(127, 127, 127); //调亮度, 实际中不能这么用 imshow(format("laplaian_layer_%d", i), dst); } else { pyrUp(pyramid_images[i], dst, pyramid_images[i - 1].size()); subtract(pyramid_images[i - 1], dst, dst);//圖像相減 dst = dst + Scalar(127, 127, 127); //調亮度, 实际中不能这么用 imshow(format("laplaian_layer_%d", i), dst); } } } void pyramid_down(vector<Mat> &pyramid_images)//高斯金字塔01 { for (int i = pyramid_images.size() - 1; i > -1; --i) { Mat dst; /* pyrUp(tmp, dst, Size(tmp.cols * 2, tmp.rows * 2)) tmp: 當前影象, 初始化為原影象 src 。 dst : 目的影象(顯示影象,為輸入影象的兩倍) Size(tmp.cols * 2, tmp.rows * 2) : 目的影象大小, 既然我們是向上取樣, pyrUp 期待一個兩倍於輸入影象(tmp)的大小。 */ pyrUp(pyramid_images[i], dst); imshow(format("pyramid_down_%d", i), dst); } } void pyramid_up(Mat &image, vector<Mat> &pyramid_images, int level)//高斯金字塔02 { Mat temp = image.clone(); Mat dst; for (int i = 0; i < level; ++i) { /* pyrDown( tmp, dst, Size( tmp.cols/2, tmp.rows/2 )) tmp: 當前影象, 初始化為原影象 src 。 dst: 目的影象( 顯示影象,為輸入影象的一半) Size( tmp.cols/2, tmp.rows/2 ) :目的影象大小, 既然我們是向下取樣, pyrDown 期待一個一半於輸入影象( tmp)的大小。 注意輸入影象的大小(在兩個方向)必須是2的冥,否則,將會顯示錯誤。 最後,將輸入影象 tmp 更新為當前顯示影象, 這樣後續操作將作用於更新後的影象。 tmp = dst; */ pyrDown(temp, dst); imshow(format("pyramid_up_%d", i), dst); temp = dst.clone(); pyramid_images.push_back(temp); } } void USMImage(Mat src, Mat &usm, float fltPar)//圖像銳化增强演算法(USM) { Mat blur_img; /* USM銳化公式表示如下: (源圖像– w*高斯模糊)/(1-w);其中w表示權重(0.1~0.9),默認為0.6 OpenCV中的代碼實現步驟 – 高斯模糊 – 權重疊加 – 輸出結果 */ GaussianBlur(src, blur_img, Size(0, 0), 25); addWeighted(src, (1 + fltPar), blur_img, (fltPar*-1), 0, usm);//原圖 : 模糊圖片= 1.5 : -0.5 的比例進行混合 imshow("usm", usm); showHistogram(usm, "Histogram_input_usm"); } void blur_demo(Mat &image, Mat &sum) { int w = image.cols; int h = image.rows; Mat result = Mat::zeros(image.size(), image.type()); int x2 = 0, y2 = 0; int x1 = 0, y1 = 0; int ksize = 5; int radius = ksize / 2; int ch = image.channels(); int cx = 0, cy = 0; for (int row = 0; row < h + radius; row++) { y2 = (row + 1)>h ? h : (row + 1); y1 = (row - ksize) < 0 ? 0 : (row - ksize); for (int col = 0; col < w + radius; col++) { x2 = (col + 1)>w ? w : (col + 1); x1 = (col - ksize) < 0 ? 0 : (col - ksize); cx = (col - radius) < 0 ? 0 : col - radius; cy = (row - radius) < 0 ? 0 : row - radius; int num = (x2 - x1)*(y2 - y1); for (int i = 0; i < ch; i++) { // 积分图查找和表,计算卷积 int s = getblockSum(sum, x1, y1, x2, y2, i); result.at<Vec3b>(cy, cx)[i] = saturate_cast<uchar>(s / num); } } } imshow("blur_demo", result); } /** * 3x3 sobel 垂直边缘检测演示 */ void edge_demo(Mat &image, Mat &sum) { int w = image.cols; int h = image.rows; Mat result = Mat::zeros(image.size(), CV_32SC3); int x2 = 0, y2 = 0; int x1 = 0, y1 = 0; int ksize = 3; // 算子大小,可以修改,越大边缘效应越明显 int radius = ksize / 2; int ch = image.channels(); int cx = 0, cy = 0; for (int row = 0; row < h + radius; row++) { y2 = (row + 1)>h ? h : (row + 1); y1 = (row - ksize) < 0 ? 0 : (row - ksize); for (int col = 0; col < w + radius; col++) { x2 = (col + 1)>w ? w : (col + 1); x1 = (col - ksize) < 0 ? 0 : (col - ksize); cx = (col - radius) < 0 ? 0 : col - radius; cy = (row - radius) < 0 ? 0 : row - radius; int num = (x2 - x1)*(y2 - y1); for (int i = 0; i < ch; i++) { // 积分图查找和表,计算卷积 int s1 = getblockSum(sum, x1, y1, cx, y2, i); int s2 = getblockSum(sum, cx, y1, x2, y2, i); result.at<Vec3i>(cy, cx)[i] = saturate_cast<int>(s2 - s1); } } } Mat dst, gray; convertScaleAbs(result, dst); normalize(dst, dst, 0, 255, NORM_MINMAX); cvtColor(dst, gray, COLOR_BGR2GRAY); imshow("edge_demo", gray); } int getblockSum(Mat &sum, int x1, int y1, int x2, int y2, int i) { int tl = sum.at<Vec3i>(y1, x1)[i]; int tr = sum.at<Vec3i>(y2, x1)[i]; int bl = sum.at<Vec3i>(y1, x2)[i]; int br = sum.at<Vec3i>(y2, x2)[i]; int s = (br - bl - tr + tl); return s; } void add_gaussian_noise(Mat &image)//高斯雜訊 { Mat noise = Mat::zeros(image.size(), image.type()); // 产生高斯噪声 randn(noise, (15, 15, 15), (30, 30, 30)); Mat dst; add(image, noise, dst); image = dst.clone();//dst.copyTo(image);//圖像複製 //imshow("gaussian_noise", dst); } void add_salt_pepper_noise(Mat &image)//白雜訊 { // 随机数产生器 RNG rng(12345); for (int i = 0; i < 1000; ++i) { int x = rng.uniform(0, image.rows); int y = rng.uniform(0, image.cols); if (i % 2 == 1) { image.at<Vec3b>(y, x) = Vec3b(255, 255, 255); } else { image.at<Vec3b>(y, x) = Vec3b(0, 0, 0); } } //imshow("saltp_epper", image); } void blur3x3(Mat &src, Mat *det) { // 3x3 均值模糊,自定义版本实现 for (int row = 1; row < src.rows - 1; row++) { for (int col = 1; col < src.cols - 1; col++) { Vec3b p1 = src.at<Vec3b>(row - 1, col - 1); Vec3b p2 = src.at<Vec3b>(row - 1, col); Vec3b p3 = src.at<Vec3b>(row - 1, col + 1); Vec3b p4 = src.at<Vec3b>(row, col - 1); Vec3b p5 = src.at<Vec3b>(row, col); Vec3b p6 = src.at<Vec3b>(row, col + 1); Vec3b p7 = src.at<Vec3b>(row + 1, col - 1); Vec3b p8 = src.at<Vec3b>(row + 1, col); Vec3b p9 = src.at<Vec3b>(row + 1, col + 1); int b = p1[0] + p2[0] + p3[0] + p4[0] + p5[0] + p6[0] + p7[0] + p8[0] + p9[0]; int g = p1[1] + p2[1] + p3[1] + p4[1] + p5[1] + p6[1] + p7[1] + p8[1] + p9[1]; int r = p1[2] + p2[2] + p3[2] + p4[2] + p5[2] + p6[2] + p7[2] + p8[2] + p9[2]; det->at<Vec3b>(row, col)[0] = saturate_cast<uchar>(b / 9); det->at<Vec3b>(row, col)[1] = saturate_cast<uchar>(g / 9); det->at<Vec3b>(row, col)[2] = saturate_cast<uchar>(r / 9); } } } void backProjection_demo(Mat &image, Mat &model)//反向投影 { Mat image_hsv, model_hsv; cvtColor(image, image_hsv, COLOR_BGR2HSV);//彩色轉HSV cvtColor(model, model_hsv, COLOR_BGR2HSV); // 定义直方图参数与属性 int h_bins = 32, s_bins = 32; int histSize[] = { h_bins, s_bins };//要切分的像素強度值範圍,預設為256。每個channel皆可指定一個範圍。例如,[32,32,32] 表示RGB三個channels皆切分為32區段 float h_ranges[] = { 0, 180 }, s_ranges[] = { 0, 256 }; const float* ranges[] = { h_ranges, s_ranges }; int channels[] = { 0, 1 }; Mat roiHist;//計算ROI的直方圖 calcHist(&model_hsv, 1, channels, Mat(), roiHist, 2, histSize, ranges); normalize(roiHist, roiHist, 0, 255, NORM_MINMAX, -1, Mat()); Mat roiproj, backproj; calcBackProject(&image_hsv, 1, channels, roiHist, roiproj, ranges);//使用反向投影 產生ROI(前景)的mask bitwise_not(roiproj, backproj);//產生背景的mask imshow("ROIProj", roiproj); imshow("BackProj", backproj); } void showHistogram(InputArray src, cv::String StrTitle)//直方圖 { bool blnGray = false; if (src.channels() == 1) { blnGray = true; } // 三通道/單通道 直方圖 紀錄陣列 vector<Mat> bgr_plane; vector<Mat> gray_plane; // 定义参数变量 const int channels[1] = { 0 }; const int bins[1] = { 256 }; float hranges[2] = { 0, 255 }; const float *ranges[1] = { hranges }; Mat b_hist, g_hist, r_hist, hist; // 计算三通道直方图 /* void calcHist( const Mat* images, int nimages,const int* channels, InputArray mask,OutputArray hist, int dims, const int* histSize,const float** ranges, bool uniform=true, bool accumulate=false ); 1.輸入的圖像數組 2.輸入數組的個數 3.通道數 4.掩碼 5.直方圖 6.直方圖維度 7.直方圖每個維度的尺寸數組 8.每一維數組的範圍 9.直方圖是否是均勻 10.配置階段不清零 */ if (blnGray) { split(src, gray_plane); calcHist(&gray_plane[0], 1, 0, Mat(), hist, 1, bins, ranges); } else { split(src, bgr_plane); calcHist(&bgr_plane[0], 1, 0, Mat(), b_hist, 1, bins, ranges); calcHist(&bgr_plane[1], 1, 0, Mat(), g_hist, 1, bins, ranges); calcHist(&bgr_plane[2], 1, 0, Mat(), r_hist, 1, bins, ranges); } /* * 显示直方图 */ int hist_w = 512; int hist_h = 400; int bin_w = cvRound((double)hist_w / bins[0]); Mat histImage = Mat::zeros(hist_h, hist_w, CV_8UC3); // 归一化直方图数据 if (blnGray) { normalize(hist, hist, 0, histImage.rows, NORM_MINMAX, -1); } else { normalize(b_hist, b_hist, 0, histImage.rows, NORM_MINMAX, -1); normalize(g_hist, g_hist, 0, histImage.rows, NORM_MINMAX, -1); normalize(r_hist, r_hist, 0, histImage.rows, NORM_MINMAX, -1); } // 绘制直方图曲线 for (int i = 1; i < bins[0]; ++i) { if (blnGray) { line(histImage, Point(bin_w * (i - 1), hist_h - cvRound(hist.at<float>(i - 1))), Point(bin_w * (i), hist_h - cvRound(hist.at<float>(i))), Scalar(255, 255, 255), 2, 8, 0); } else { line(histImage, Point(bin_w * (i - 1), hist_h - cvRound(b_hist.at<float>(i - 1))), Point(bin_w * (i), hist_h - cvRound(b_hist.at<float>(i))), Scalar(255, 0, 0), 2, 8, 0); line(histImage, Point(bin_w * (i - 1), hist_h - cvRound(g_hist.at<float>(i - 1))), Point(bin_w * (i), hist_h - cvRound(g_hist.at<float>(i))), Scalar(0, 255, 0), 2, 8, 0); line(histImage, Point(bin_w * (i - 1), hist_h - cvRound(r_hist.at<float>(i - 1))), Point(bin_w * (i), hist_h - cvRound(r_hist.at<float>(i))), Scalar(0, 0, 255), 2, 8, 0); } } imshow(StrTitle, histImage); }
★Python
import cv2 as cv import numpy as np # # THRESH_BINARY = 0 # THRESH_BINARY_INV = 1 # THRESH_TRUNC = 2 # THRESH_TOZERO = 3 # THRESH_TOZERO_INV = 4 # src = cv.imread("D:/images/lena.jpg") cv.namedWindow("input", cv.WINDOW_AUTOSIZE) cv.imshow("input", src) h, w = src.shape[:2] # 自动阈值分割 OTSU gray = cv.cvtColor(src, cv.COLOR_BGR2GRAY) ret, binary = cv.threshold(gray, 0, 255, cv.THRESH_BINARY | cv.THRESH_OTSU) print("ret :", ret) cv.imshow("binary", binary) result = np.zeros([h, w*2, 3], dtype=src.dtype) result[0:h,0:w,:] = src result[0:h,w:2*w,:] = cv.cvtColor(binary, cv.COLOR_GRAY2BGR) cv.putText(result, "input", (10, 30), cv.FONT_ITALIC, 1.0, (0, 0, 255), 2) cv.putText(result, "binary, threshold = " + str(ret), (w+10, 30), cv.FONT_ITALIC, 1.0, (0, 0, 255), 2) cv.imshow("result", result) cv.imwrite("D:/binary_result.png", result) cv.waitKey(0) cv.destroyAllWindows()
★結果圖:
★延伸說明/重點回顧: