【OpenCV】Chapter9.边缘检测与图像分割

2022-09-28 20:48:07 浏览数 (1)

最近想对OpenCV进行系统学习,看到网上这份教程写得不错,于是跟着来学习实践一下。 【youcans@qq.com, youcans 的 OpenCV 例程, https://youcans.blog.csdn.net/article/details/125112487】 程序仓库:https://github.com/zstar1003/OpenCV-Learning

边缘检测

Roberts算子/Prewitt算子/Sobel算子/Laplacian算子

边缘检测的原理和matlab实现在我之前这两篇博文中提到过,这里不再赘述。

【计算机视觉】基础图像知识点整理【计算机视觉】数字图像处理基础知识题

此次来看OpenCV的实现方式。

OpenCV并没有直接提供相应的函数接口,因此通过自定义卷积核可以实现各种边缘检测算子。

示例程序:

代码语言:javascript复制
"""
边缘检测(Roberts算子, Prewitt算子, Sobel算子, Laplacian算子)
"""
import cv2
import matplotlib.pyplot as plt
import numpy as np

img = cv2.imread("../img/lena.jpg", flags=0)

# 自定义卷积核
# Roberts 边缘算子
kernel_Roberts_x = np.array([[1, 0], [0, -1]])
kernel_Roberts_y = np.array([[0, -1], [1, 0]])
# Prewitt 边缘算子
kernel_Prewitt_x = np.array([[-1, 0, 1], [-1, 0, 1], [-1, 0, 1]])
kernel_Prewitt_y = np.array([[1, 1, 1], [0, 0, 0], [-1, -1, -1]])
# Sobel 边缘算子
kernel_Sobel_x = np.array([[-1, 0, 1], [-2, 0, 2], [-1, 0, 1]])
kernel_Sobel_y = np.array([[1, 2, 1], [0, 0, 0], [-1, -2, -1]])
# Laplacian 边缘算子
kernel_Laplacian_K1 = np.array([[0, 1, 0], [1, -4, 1], [0, 1, 0]])
kernel_Laplacian_K2 = np.array([[1, 1, 1], [1, -8, 1], [1, 1, 1]])

# 卷积运算
imgBlur = cv2.blur(img, (3, 3))  # Blur 平滑后再做 Laplacian 变换
imgLaplacian_K1 = cv2.filter2D(imgBlur, -1, kernel_Laplacian_K1)
imgLaplacian_K2 = cv2.filter2D(imgBlur, -1, kernel_Laplacian_K2)
imgRoberts_x = cv2.filter2D(img, -1, kernel_Roberts_x)
imgRoberts_y = cv2.filter2D(img, -1, kernel_Roberts_y)
imgRoberts = np.uint8(cv2.normalize(abs(imgRoberts_x)   abs(imgRoberts_y), None, 0, 255, cv2.NORM_MINMAX))
imgPrewitt_x = cv2.filter2D(img, -1, kernel_Prewitt_x)
imgPrewitt_y = cv2.filter2D(img, -1, kernel_Prewitt_y)
imgPrewitt = np.uint8(cv2.normalize(abs(imgPrewitt_x)   abs(imgPrewitt_y), None, 0, 255, cv2.NORM_MINMAX))
imgSobel_x = cv2.filter2D(img, -1, kernel_Sobel_x)
imgSobel_y = cv2.filter2D(img, -1, kernel_Sobel_y)
imgSobel = np.uint8(cv2.normalize(abs(imgSobel_x)   abs(imgSobel_y), None, 0, 255, cv2.NORM_MINMAX))

plt.figure(figsize=(12, 8))
plt.subplot(341), plt.title('Origin'), plt.imshow(img, cmap='gray'), plt.axis('off')
plt.subplot(345), plt.title('Laplacian_K1'), plt.imshow(imgLaplacian_K1, cmap='gray'), plt.axis('off')
plt.subplot(349), plt.title('Laplacian_K2'), plt.imshow(imgLaplacian_K2, cmap='gray'), plt.axis('off')
plt.subplot(342), plt.title('Roberts'), plt.imshow(imgRoberts, cmap='gray'), plt.axis('off')
plt.subplot(346), plt.title('Roberts_X'), plt.imshow(imgRoberts_x, cmap='gray'), plt.axis('off')
plt.subplot(3, 4, 10), plt.title('Roberts_Y'), plt.imshow(imgRoberts_y, cmap='gray'), plt.axis('off')
plt.subplot(343), plt.title('Prewitt'), plt.imshow(imgPrewitt, cmap='gray'), plt.axis('off')
plt.subplot(347), plt.title('Prewitt_X'), plt.imshow(imgPrewitt_x, cmap='gray'), plt.axis('off')
plt.subplot(3, 4, 11), plt.title('Prewitt_Y'), plt.imshow(imgPrewitt_y, cmap='gray'), plt.axis('off')
plt.subplot(344), plt.title('Sobel'), plt.imshow(imgSobel, cmap='gray'), plt.axis('off')
plt.subplot(348), plt.title('Sobel_X'), plt.imshow(imgSobel_x, cmap='gray'), plt.axis('off')
plt.subplot(3, 4, 12), plt.title('Sobel_Y'), plt.imshow(imgSobel_y, cmap='gray'), plt.axis('off')
plt.tight_layout()
plt.show()

LoG算子(Marr-Hildreth 算法)

Marr-Hildreth 算法是改进的边缘检测算子,是平滑算子与 Laplace 算子的结合,因而兼具平滑和二阶微分的作用,其定位精度高,边缘连续性好,计算速度快。

示例程序:

代码语言:javascript复制
"""
LoG边缘检测算子
"""
import cv2
import matplotlib.pyplot as plt
import numpy as np
from scipy import signal

img = cv2.imread("../img/lena.jpg", flags=0)


def ZeroDetect(img):  # 判断零交叉点
    h, w = img.shape[0], img.shape[1]
    zeroCrossing = np.zeros_like(img, np.uint8)
    for x in range(0, w - 1):
        for y in range(0, h - 1):
            if img[y][x] < 0:
                if (img[y][x - 1] > 0) or (img[y][x   1] > 0) 
                        or (img[y - 1][x] > 0) or (img[y   1][x] > 0):
                    zeroCrossing[y][x] = 255
    return zeroCrossing

imgBlur = cv2.blur(img, (3, 3))  # Blur 平滑后再做 Laplacian 变换

# 近似的 Marr-Hildreth 卷积核 (5*5)
kernel_MH5 = np.array([
    [0, 0, -1, 0, 0],
    [0, -1, -2, -1, 0],
    [-1, -2, 16, -2, -1],
    [0, -1, -2, -1, 0],
    [0, 0, -1, 0, 0]])
imgMH5 = signal.convolve2d(imgBlur, kernel_MH5, boundary='symm', mode='same')  # 卷积计算
zeroMH5 = ZeroDetect(imgMH5)  # 判断零交叉点

# 由 Gauss 标准差计算 Marr-Hildreth 卷积核
sigma = 3  # Gauss 标准差,输入参数
size = int(2 * round(3 * sigma))   1  # 根据标准差确定窗口大小,3*sigma 占比 99.7%
print("sigma={:d}, size={}".format(sigma, size))
x, y = np.meshgrid(np.arange(-size / 2   1, size / 2   1), np.arange(-size / 2   1, size / 2   1))  # 生成网格
norm2 = np.power(x, 2)   np.power(y, 2)
sigma2, sigma4 = np.power(sigma, 2), np.power(sigma, 4)
kernelLoG = ((norm2 - (2.0 * sigma2)) / sigma4) * np.exp(- norm2 / (2.0 * sigma2))  # 计算 LoG 卷积核
# Marr-Hildreth 卷积运算
imgLoG = signal.convolve2d(imgBlur, kernelLoG, boundary='symm', mode='same')  # 卷积计算
# 判断零交叉点
zeroCrossing = ZeroDetect(imgLoG)

plt.figure(figsize=(10, 7))
plt.subplot(221), plt.title("Marr-Hildreth (sigma=0.5)"), plt.imshow(imgMH5, cmap='gray'), plt.axis('off')
plt.subplot(222), plt.title("Marr-Hildreth (sigma=3)"), plt.imshow(imgLoG, cmap='gray'), plt.axis('off')
plt.subplot(223), plt.title("Zero crossing (size=5)"), plt.imshow(zeroMH5, cmap='gray'), plt.axis('off')
plt.subplot(224), plt.title("Zero crossing (size=19)"), plt.imshow(zeroCrossing, cmap='gray'), plt.axis('off')
plt.tight_layout()
plt.show()

DoG算子

DoG算子是对LoG算子的简化。

示例程序:

代码语言:javascript复制
"""
DoG边缘检测算子
"""
import cv2
import matplotlib.pyplot as plt
import numpy as np
from scipy import signal

img = cv2.imread("../img/lena.jpg", flags=0)


# 高斯核低通滤波器,sigmaY 缺省时 sigmaY=sigmaX
kSize = (5, 5)
imgGaussBlur1 = cv2.GaussianBlur(img, (5, 5), sigmaX=1.0)  # sigma=1.0
imgGaussBlur2 = cv2.GaussianBlur(img, (5, 5), sigmaX=2.0)  # sigma=2.0
imgGaussBlur3 = cv2.GaussianBlur(img, (5, 5), sigmaX=4.0)  # sigma=4.0
imgGaussBlur4 = cv2.GaussianBlur(img, (5, 5), sigmaX=16.0)  # sigma=16.0

# 高斯差分算子 (Difference of Gaussian)
imgDoG1 = imgGaussBlur2 - imgGaussBlur1  # sigma=1.0,2.0
imgDoG2 = imgGaussBlur3 - imgGaussBlur2  # sigma=2.0,4.0
imgDoG3 = imgGaussBlur4 - imgGaussBlur3  # sigma=4.0,16.0

plt.figure(figsize=(10, 6))
plt.subplot(231), plt.title("GaussBlur (sigma=2.0)"), plt.imshow(imgGaussBlur2, cmap='gray'), plt.axis('off')
plt.subplot(232), plt.title("GaussBlur (sigma=4.0)"), plt.imshow(imgGaussBlur3, cmap='gray'), plt.axis('off')
plt.subplot(233), plt.title("GaussBlur (sigma=16.)"), plt.imshow(imgGaussBlur4, cmap='gray'), plt.axis('off')
plt.subplot(234), plt.title("DoG (sigma=1.0,2.0)"), plt.imshow(imgDoG1, cmap='gray'), plt.axis('off')
plt.subplot(235), plt.title("DoG (sigma=2.0,4.0)"), plt.imshow(imgDoG2, cmap='gray'), plt.axis('off')
plt.subplot(236), plt.title("DoG (sigma=4.0,16.)"), plt.imshow(imgDoG3, cmap='gray'), plt.axis('off')
plt.tight_layout()
plt.show()

Canny算子

Canny算子执行的基本步骤为:

(1)使用高斯滤波对图像进行平滑;

(2)用一阶有限差分计算梯度幅值和方向;

(3)对梯度幅值进行非极大值抑制(NMS);

(4)用双阈值处理和连通性分析来检测和连接边缘

OpenCV提供了函数cv.Canny实现Canny边缘检测算子。

cv.Canny(image, threshold1, threshold2[, edges[, apertureSize, L2gradient]]) → edges

参数说明:

  • image:输入图像,8-bit 灰度图像,不适用彩色图像
  • edges:输出边缘图像,8-bit 单通道图像,大小与输入图像相同
  • threshold1:第一阈值 TL
  • threshold2:第二阈值 TH
  • apertureSize:Sobel 卷积核的孔径,可选项,默认值 3
  • L2gradient: 计算图像梯度幅值 标志符,默认值为 True 表示 L2 法,False 表示 L1 法

示例程序:

代码语言:javascript复制
"""
Canny边缘检测算子
"""
import cv2
import matplotlib.pyplot as plt
import numpy as np

img = cv2.imread("../img/lena.jpg", flags=0)

# 高斯核低通滤波器,sigmaY 缺省时 sigmaY=sigmaX
kSize = (5, 5)
imgGauss1 = cv2.GaussianBlur(img, kSize, sigmaX=1.0)  # sigma=1.0
imgGauss2 = cv2.GaussianBlur(img, kSize, sigmaX=10.0)  # sigma=2.0

# 高斯差分算子 (Difference of Gaussian)
imgDoG = imgGauss2 - imgGauss1  # sigma=1.0, 10.0

# Canny 边缘检测, kSize 为高斯核大小,t1,t2为阈值大小
t1, t2 = 50, 150
imgCanny = cv2.Canny(imgGauss1, t1, t2)

plt.figure(figsize=(10, 6))
plt.subplot(131), plt.title("Origin"), plt.imshow(img, cmap='gray'), plt.axis('off')
plt.subplot(132), plt.title("DoG"), plt.imshow(imgDoG, cmap='gray'), plt.axis('off')
plt.subplot(133), plt.title("Canny"), plt.imshow(imgCanny, cmap='gray'), plt.axis('off')
plt.tight_layout()
plt.show()

图像分割

区域生长

区域生长方法将具有相似性质的像素或子区域组合为更大区域。

区域增长方法的步骤:

(1)对图像自上而下、从左向右扫描,找到第 1 个还没有访问过的像素,将该像素作为种子 (x0, y0);

(2)以 (x0, y0) 为中心, 考虑其 4 邻域或 8 邻域像素 (x, y),如果其邻域满足生长准则 则将 (x, y) 与 (x0, y0) 合并到同一区域,同时将 (x, y) 压入堆栈;

(3)从堆栈中取出一个像素,作为种子 (x0, y0) 继续步骤(2);

(4)当堆栈为空时返回步骤(1);

(5)重复步骤(1)-(4),直到图像中的每个点都被访问过,算法结束

示例程序:

代码语言:javascript复制
"""
图像分割之区域生长
"""
import cv2
import matplotlib.pyplot as plt
import numpy as np


def getGrayDiff(image, currentPoint, tmpPoint):  # 求两个像素的距离
    return abs(int(image[currentPoint[0], currentPoint[1]]) - int(image[tmpPoint[0], tmpPoint[1]]))


# 区域生长算法
def regional_growth(img, seeds, thresh=5):
    height, weight = img.shape
    seedMark = np.zeros(img.shape)
    seedList = []
    for seed in seeds:
        if (0 < seed[0] < height and 0 < seed[1] < weight): seedList.append(seed)
    label = 1  # 种子位置标记
    connects = [(-1, -1), (0, -1), (1, -1), (1, 0), (1, 1), (0, 1), (-1, 1), (-1, 0)]  # 8 邻接连通
    while (len(seedList) > 0):  # 如果列表里还存在点
        currentPoint = seedList.pop(0)  # 将最前面的那个抛出
        seedMark[currentPoint[0], currentPoint[1]] = label  # 将对应位置的点标记为 1
        for i in range(8):  # 对这个点周围的8个点一次进行相似性判断
            tmpX = currentPoint[0]   connects[i][0]
            tmpY = currentPoint[1]   connects[i][1]
            if tmpX < 0 or tmpY < 0 or tmpX >= height or tmpY >= weight:  # 是否超出限定阈值
                continue
            grayDiff = getGrayDiff(img, currentPoint, (tmpX, tmpY))  # 计算灰度差
            if grayDiff < thresh and seedMark[tmpX, tmpY] == 0:
                seedMark[tmpX, tmpY] = label
                seedList.append((tmpX, tmpY))
    return seedMark


img = cv2.imread("../img/lena.jpg", flags=0)
# histCV = cv2.calcHist([img], [0], None, [256], [0, 256])  # 灰度直方图
# OTSU 全局阈值处理
ret, imgOtsu = cv2.threshold(img, 127, 255, cv2.THRESH_OTSU)  # 阈值分割, thresh=T
# 自适应局部阈值处理
binaryMean = cv2.adaptiveThreshold(img, 255, cv2.ADAPTIVE_THRESH_MEAN_C, cv2.THRESH_BINARY, 5, 3)
# 区域生长图像分割
# seeds = [(10, 10), (82, 150), (20, 300)]  # 直接给定 种子点
imgBlur = cv2.blur(img, (3, 3))  # cv2.blur 方法
_, imgTop = cv2.threshold(imgBlur, 250, 255, cv2.THRESH_BINARY)  # 高百分位阈值产生种子区域
nseeds, labels, stats, centroids = cv2.connectedComponentsWithStats(imgTop)  # 过滤连通域,获得质心点 (x,y)
seeds = centroids.astype(int)  # 获得质心像素作为种子点
imgGrowth = regional_growth(img, seeds, 8)

plt.figure(figsize=(8, 6))
plt.subplot(221), plt.axis('off'), plt.title("Origin")
plt.imshow(img, 'gray')
plt.subplot(222), plt.axis('off'), plt.title("OTSU(T={})".format(ret))
plt.imshow(imgOtsu, 'gray')
plt.subplot(223), plt.axis('off'), plt.title("Adaptive threshold")
plt.imshow(binaryMean, 'gray')
plt.subplot(224), plt.axis('off'), plt.title("Region grow")
plt.imshow(255 - imgGrowth, 'gray')
plt.tight_layout()
plt.show()

区域分离

区域分离的判据是用户选择的谓词逻辑Q,通常是目标区域特征一致性的测度,例如灰度均值和方差。

分离过程先判断当前区域是否满足目标的特征测度,如果不满足则将当前区域分离为多个子区域进行判断;不断重复判断、分离,直到拆分到最小区域为止。

示例程序:

代码语言:javascript复制
"""
图像分割之区域分离
"""
import cv2
import matplotlib.pyplot as plt
import numpy as np


def SplitMerge(src, dst, h, w, h0, w0, maxMean, minVar, cell=4):
    win = src[h0: h0   h, w0: w0   w]
    mean = np.mean(win)  # 窗口区域的均值
    var = np.std(win, ddof=1)  # 窗口区域的标准差,无偏样本标准差

    if (mean < maxMean) and (var > minVar) and (h < 2 * cell) and (w < 2 * cell):
        # 该区域满足谓词逻辑条件,判为目标区域,设为白色
        dst[h0:h0   h, w0:w0   w] = 255  # 白色
        # print("h0={}, w0={}, h={}, w={}, mean={:.2f}, var={:.2f}".
        #       format(h0, w0, h, w, mean, var))
    else:  # 该区域不满足谓词逻辑条件
        if (h > cell) and (w > cell):  # 区域能否继续分拆?继续拆
            SplitMerge(src, dst, (h   1) // 2, (w   1) // 2, h0, w0, maxMean, minVar, cell)
            SplitMerge(src, dst, (h   1) // 2, (w   1) // 2, h0, w0   (w   1) // 2, maxMean, minVar, cell)
            SplitMerge(src, dst, (h   1) // 2, (w   1) // 2, h0   (h   1) // 2, w0, maxMean, minVar, cell)
            SplitMerge(src, dst, (h   1) // 2, (w   1) // 2, h0   (h   1) // 2, w0   (w   1) // 2, maxMean, minVar,
                       cell)
        # else:  # 不能再分拆,判为非目标区域,设为黑色
        #     src[h0:h0 h, w0:w0 w] = 0  # 黑色


img = cv2.imread("../img/lena.jpg", flags=0)
hImg, wImg = img.shape
mean = np.mean(img)  # 窗口区域的均值
var = np.std(img, ddof=1)  # 窗口区域的标准差,无偏样本标准差
print("h={}, w={}, mean={:.2f}, var={:.2f}".format(hImg, wImg, mean, var))

maxMean = 80  # 均值上界
minVar = 10  # 标准差下界
src = img.copy()
dst1 = np.zeros_like(img)
dst2 = np.zeros_like(img)
dst3 = np.zeros_like(img)
SplitMerge(src, dst1, hImg, wImg, 0, 0, maxMean, minVar, cell=32)  # 最小分割区域 cell=32
SplitMerge(src, dst2, hImg, wImg, 0, 0, maxMean, minVar, cell=16)  # 最小分割区域 cell=16
SplitMerge(src, dst3, hImg, wImg, 0, 0, maxMean, minVar, cell=8)  # 最小分割区域 cell=8

plt.figure(figsize=(9, 7))
plt.subplot(221), plt.axis('off'), plt.title("Origin")
plt.imshow(img, 'gray')
plt.subplot(222), plt.axis('off'), plt.title("Region split (c=32)")
plt.imshow(dst1, 'gray')
plt.subplot(223), plt.axis('off'), plt.title("Region split (c=16)")
plt.imshow(dst2, 'gray')
plt.subplot(224), plt.axis('off'), plt.title("Region split (c=8)")
plt.imshow(dst3, 'gray')
plt.tight_layout()
plt.show()

K均值聚类

OpenCV 提供了函数cv.kmeans来实现 k-means 聚类算法。

cv.kmeans(data, K, bestLabels, criteria, attempts, flags, centers) → compactness, labels, centersdst

参数说明:

  • data:用于聚类的数据,N 维数组,类型为 CV_32F、CV_32FC2
  • K:设定的聚类数量
  • bestLabels:整数数组,分类标签,每个样本的所属聚类的序号
  • criteria:元组 (type, max_iter, epsilon),算法结束标准,最大迭代次数或聚类中心位置精度
    • cv2.TERM_CRITERIA_EPS:如果达到指定的精度 epsilon,则停止算法迭代
    • cv2.TERM_CRITERIA_MAX_ITER:在指定的迭代次数max_iter之后停止算法
    • cv2.TERM_CRITERIA_EPS cv.TERM_CRITERIA_MAX_ITER:当满足上述任何条件时停止迭代
  • attempts:标志,指定使用不同聚类中心初值执行算法的次数
  • flags:像素邻域的尺寸,用于计算邻域的阈值,通常取 3,5,7
    • cv2. KMEANS_RANDOM_CENTERS:随机产生聚类中心的初值
    • cv2. KMEANS_PP_CENTERS:Kmeans 中心初始化方法
    • cv2. KMEANS_USE_INITIAL_LABELS:第一次计算时使用用户指定的聚类初值,之后的计算则使用随机的或半随机的聚类中心初值
  • centers:聚类中心数组,每个聚类中心为一行,可选项
  • labels:整数数组,分类标签,每个样本的所属聚类的序号
  • centersdst:聚类中心数组

示例程序:

代码语言:javascript复制
"""
图像分割之k均值聚类
"""
import cv2
import matplotlib.pyplot as plt
import numpy as np

img = cv2.imread("../img/img.jpg", flags=1)

dataPixel = np.float32(img.reshape((-1, 3)))
criteria = (cv2.TERM_CRITERIA_EPS   cv2.TERM_CRITERIA_MAX_ITER, 200, 0.1)  # 终止条件
flags = cv2.KMEANS_RANDOM_CENTERS  # 起始的中心选择

K = 2  # 设置聚类数
_, labels, center = cv2.kmeans(dataPixel, K, None, criteria, 10, flags)
centerUint = np.uint8(center)
classify = centerUint[labels.flatten()]  # 将像素标记为聚类中心颜色
imgKmean3 = classify.reshape((img.shape))  # 恢复为二维图像

K = 3  # 设置聚类数
_, labels, center = cv2.kmeans(dataPixel, K, None, criteria, 10, flags)
centerUint = np.uint8(center)
classify = centerUint[labels.flatten()]  # 将像素标记为聚类中心颜色
imgKmean4 = classify.reshape((img.shape))  # 恢复为二维图像

K = 5  # 设置聚类数
_, labels, center = cv2.kmeans(dataPixel, K, None, criteria, 10, flags)
centerUint = np.uint8(center)
classify = centerUint[labels.flatten()]  # 将像素标记为聚类中心颜色
imgKmean5 = classify.reshape((img.shape))  # 恢复为二维图像

plt.figure(figsize=(9, 7))
plt.subplot(221), plt.axis('off'), plt.title("Origin")
plt.imshow(cv2.cvtColor(img, cv2.COLOR_BGR2RGB))  # 显示 img1(RGB)
plt.subplot(222), plt.axis('off'), plt.title("K-mean (k=2)")
plt.imshow(cv2.cvtColor(imgKmean3, cv2.COLOR_BGR2RGB))
plt.subplot(223), plt.axis('off'), plt.title("K-mean (k=3)")
plt.imshow(cv2.cvtColor(imgKmean4, cv2.COLOR_BGR2RGB))
plt.subplot(224), plt.axis('off'), plt.title("K-mean (k=5)")
plt.imshow(cv2.cvtColor(imgKmean5, cv2.COLOR_BGR2RGB))
plt.tight_layout()
plt.show()

寻找轮廓

OpenCV提供函数cv.findContours()从二值图像中寻找轮廓,函数cv2.drawContours()绘制轮廓。

cv.findContours(image, mode, method[, contours[, hierarchy, offset]]) → contours, hierarchy

参数说明:

  • image:原始图像,8 位单通道二值图像
  • mode: 轮廓检索模式
    • cv.RETR_EXTERNAL:只检索最外层轮廓
    • cv.RETR_LIST:检索所有轮廓,不建立任何层次关系
    • cv.RETR_CCOMP:检索所有轮廓,并将其组织为两层, 顶层是各部分的外部轮廓,次层是内层轮廓
    • cv.RETR_TREE:检索所有轮廓,并重建嵌套轮廓的完整层次结构
    • cv.RETR_FLOODFILL:漫水填充法(泛洪填充)
  • method: 轮廓近似方法
    • cv.CHAIN_APPROX_NONE:输出轮廓的每个像素点
    • cv.CHAIN_APPROX_SIMPLE:压缩水平、垂直和斜线,仅保留这些线段的端点
    • cv.CHAIN_APPROX_TC89_L1:应用 Teh-Chin 链近似算法 L1
    • cv.CHAIN_APPROX_TC89_KCOS:应用 Teh-Chin 链近似算法 KCOS
  • contours:检测到的所有轮廓,列表格式,每个轮廓存储为包含边界点坐标 (x,y) 的点向量
    • 列表(LIST)长度为 L,对应于找到的 L 个轮廓,按 0,…L-1 顺序排列
    • 列表中的第 i 个元素是一个形如 (k,1,2) 的 Numpy 数组,表示第 i 个轮廓,k 是第 i 个轮廓的边界点的数量
    • 数组 contoursi 是构成第 i 个轮廓的各边界点坐标 (x,y) 的点向量
    • 注意边界点的坐标表达形式是 (x,y),而不是 OpenCV 中常用的像素坐标表达形式 (y,x)。
  • hierarchy:轮廓的层次结构和拓扑信息,是一个形如 (1,k,4) 的 Numpy 数组
    • k 对应于找到的轮廓数量
    • hierarchy0 表示第 i 个轮廓的层次结构,是包含 4个值的数组 Next, Previous, First Child, Parent,分别代表第 i 个轮廓的同层的后一个轮廓、同层的前一个轮廓、第一个子轮廓、父轮廓的编号
  • offset:每个轮廓点的偏移量

示例程序:

代码语言:javascript复制
"""
图像分割之绘制轮廓
"""
import cv2
import matplotlib.pyplot as plt
import numpy as np

img = cv2.imread("../img/img.jpg", flags=1)

gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)  # 灰度图像
_, binary = cv2.threshold(gray, 127, 255, cv2.THRESH_OTSU   cv2.THRESH_BINARY_INV)
plt.figure(figsize=(9, 6))
plt.subplot(131), plt.axis('off'), plt.title("Origin")
plt.imshow(cv2.cvtColor(img, cv2.COLOR_BGR2RGB))
plt.subplot(132), plt.axis('off'), plt.title("BinaryInv")
plt.imshow(binary, 'gray')

# 寻找二值化图中的轮廓
binary, contours, hierarchy = cv2.findContours(binary, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)  # OpenCV3
# contours, hierarchy = cv2.findContours(binary, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)  # OpenCV4~
# # 绘制轮廓
contourPic = img.copy()  # OpenCV3.2 之前的早期版本,查找轮廓函数会修改原始图像
contourPic = cv2.drawContours(contourPic, contours, -1, (0, 0, 255), 2)  # OpenCV3
# contourPic = cv.drawContours(img, contours, -1, (0, 0, 255), thickness=cv.FILLED,maxLevel=1)

print("len(contours) = ", len(contours))  # 所有轮廓的列表
for i in range(len(contours)):
    print("i=", i, contours[i].shape)  # 第 i 个轮廓的边界点
print("hierarchy.shape : ", hierarchy.shape)  # 层次结构
print(hierarchy)

plt.subplot(133), plt.axis('off'), plt.title("External contour")
plt.imshow(cv2.cvtColor(contourPic, cv2.COLOR_BGR2RGB))
plt.tight_layout()
plt.show()

趣味应用

下面展示两个趣味应用,原理就不细述了,可以根据阅读代码理解。

GraphCuts图割法

GraphCuts图割法作者为youcans,利用OpenCV实现的交互性应用。通过用鼠标左键标记前景,鼠标右键标记背景,然后实现分割。

代码:

代码语言:javascript复制
'''
GraphCuts 交互式图割分割算法

说明:
  (1) 用鼠标左键标记前景,鼠标右键标记背景;
  (2) 可以重复标记,不断优化;
  (3) 按 Esc 键退出,完成分割。
'''

import cv2
import numpy as np
from matplotlib import pyplot as plt

drawing = False
mode = False


class GraphCutXupt:
    def __init__(self, t_img):
        self.img = t_img
        self.img_raw = img.copy()
        self.img_width = img.shape[0]
        self.img_height = img.shape[1]
        self.scale_size = 640 * self.img_width // self.img_height
        if self.img_width > 640:
            self.img = cv2.resize(self.img, (640, self.scale_size), interpolation=cv2.INTER_AREA)
        self.img_show = self.img.copy()
        self.img_gc = self.img.copy()
        self.img_gc = cv2.GaussianBlur(self.img_gc, (3, 3), 0)
        self.lb_up = False
        self.rb_up = False
        self.lb_down = False
        self.rb_down = False
        self.mask = np.full(self.img.shape[:2], 2, dtype=np.uint8)
        self.firt_choose = True


# 鼠标的回调函数
def mouse_event2(event, x, y, flags, param):
    global drawing, last_point, start_point
    # 左键按下:开始画图
    if event == cv2.EVENT_LBUTTONDOWN:
        drawing = True
        last_point = (x, y)
        start_point = last_point
        param.lb_down = True
        print('mouse lb down')
    elif event == cv2.EVENT_RBUTTONDOWN:
        drawing = True
        last_point = (x, y)
        start_point = last_point
        param.rb_down = True
        print('mouse rb down')
    # 鼠标移动,画图
    elif event == cv2.EVENT_MOUSEMOVE:
        if drawing:
            if param.lb_down:
                cv2.line(param.img_show, last_point, (x, y), (0, 0, 255), 2, -1)
                cv2.rectangle(param.mask, last_point, (x, y), 1, -1, 4)
            else:
                cv2.line(param.img_show, last_point, (x, y), (255, 0, 0), 2, -1)
                cv2.rectangle(param.mask, last_point, (x, y), 0, -1, 4)
            last_point = (x, y)
    # 左键释放:结束画图
    elif event == cv2.EVENT_LBUTTONUP:
        drawing = False
        param.lb_up = True
        param.lb_down = False
        cv2.line(param.img_show, last_point, (x, y), (0, 0, 255), 2, -1)
        if param.firt_choose:
            param.firt_choose = False
        cv2.rectangle(param.mask, last_point, (x, y), 1, -1, 4)
        # print('mouse lb up')
    elif event == cv2.EVENT_RBUTTONUP:
        drawing = False
        param.rb_up = True
        param.rb_down = False
        cv2.line(param.img_show, last_point, (x, y), (255, 0, 0), 2, -1)
        if param.firt_choose:
            param.firt_choose = False
            param.mask = np.full(param.img.shape[:2], 3, dtype=np.uint8)
        cv2.rectangle(param.mask, last_point, (x, y), 0, -1, 4)
        # print('mouse rb up')


if __name__ == '__main__':
    img = cv2.imread("../img/img.jpg", flags=1)  # 读取彩色图像(Youcans)
    g_img = GraphCutXupt(img)

    cv2.namedWindow('image')
    # 定义鼠标的回调函数
    cv2.setMouseCallback('image', mouse_event2, g_img)
    while (True):
        cv2.imshow('image', g_img.img_show)
        if g_img.lb_up or g_img.rb_up:
            g_img.lb_up = False
            g_img.rb_up = False
            bgdModel = np.zeros((1, 65), np.float64)
            fgdModel = np.zeros((1, 65), np.float64)
            rect = (1, 1, g_img.img.shape[1], g_img.img.shape[0])
            # print(g_img.mask)
            mask = g_img.mask
            g_img.img_gc = g_img.img.copy()
            cv2.grabCut(g_img.img_gc, mask, rect, bgdModel, fgdModel, 5, cv2.GC_INIT_WITH_MASK)
            mask2 = np.where((mask == 2) | (mask == 0), 0, 1).astype('uint8')  # 0和2做背景
            g_img.img_gc = g_img.img_gc * mask2[:, :, np.newaxis]  # 使用蒙板来获取前景区域
            cv2.imshow('result', g_img.img_gc)
        # 按下ESC键退出
        if cv2.waitKey(20) == 27:
            break

    plt.figure(figsize=(10, 7))
    plt.subplot(221), plt.axis('off'), plt.title("xupt")
    plt.imshow(cv2.cvtColor(img, cv2.COLOR_BGR2RGB))  # 显示 img(RGB)
    plt.subplot(222), plt.axis('off'), plt.title("mask")
    plt.imshow(mask, 'gray')
    plt.subplot(223), plt.axis('off'), plt.title("mask2")
    plt.imshow(mask2, 'gray')
    plt.subplot(224), plt.axis('off'), plt.title("Grab Cut")
    plt.imshow(cv2.cvtColor(g_img.img_gc, cv2.COLOR_BGR2RGB))
    plt.tight_layout()
    plt.show()

有点类似于剪辑软件中的抠图笔刷。

旧影浮光——色彩保留滤镜

该demo的作者是ZouKeh,通过一个交互性界面,可以在原图灰度图上绘制矩形,从而框选出有色彩的部分。

程序代码:

代码语言:javascript复制
"""
Title:旧影浮光——基于Opencv的色彩保留滤镜
Author:ZouKeh
Link:https://www.bilibili.com/video/BV1BG4y1r7WP
"""

import numpy as np
import cv2


def color_choose():
    # 选择需要保留的颜色
    color = int(input("choose color(1.red 2.yellow 3.blue 4.green 5.white):"))
    lower = np.array([0, 0, 0])
    upper = np.array([0, 0, 0])
    if color == 1:
        lower = np.array([156, 60, 60])
        upper = np.array([180, 255, 255])
        # lower = np.array([0, 60, 60])
        # upper = np.array([10, 255, 255])
    elif color == 2:
        lower = np.array([26, 43, 46])
        upper = np.array([34, 255, 255])
    elif color == 3:
        lower = np.array([100, 43, 46])
        upper = np.array([124, 255, 255])
    elif color == 4:
        lower = np.array([35, 43, 46])
        upper = np.array([77, 255, 255])
    elif color == 5:
        lower = np.array([0, 0, 221])
        upper = np.array([180, 30, 255])
    return lower, upper, color


def choose_range(img):
    # 在图片上画区域 选择需要保留颜色的区域
    a = []
    b = []

    ##鼠标事件 左键单击
    def on_EVENT_LBUTTONDOWN(event, x, y, flags, param):
        global num  # 界面上点的个数 0开始 偶数为起点 奇数为终点
        # global begin_xy
        if event == cv2.EVENT_LBUTTONDOWN and num % 2 == 0:  # 画选框的左上角 红色
            xy = "%d,%d" % (x, y)
            a.append(x)
            b.append(y)
            cv2.circle(img, (x, y), 2, (0, 0, 255), thickness=-1)
            cv2.putText(img, xy, (x, y), cv2.FONT_HERSHEY_PLAIN,
                        1.0, (0, 0, 0), thickness=1)
            cv2.imshow("image", img)
            print(x, y)
            num  = 1
            begin_xy = (x, y)
        elif event == cv2.EVENT_LBUTTONDOWN and num % 2 == 1:  # 画选框的右下角 绿色
            xy = "%d,%d" % (x, y)
            a.append(x)
            b.append(y)
            cv2.circle(img, (x, y), 2, (0, 255, 0), thickness=-1)
            cv2.putText(img, xy, (x, y), cv2.FONT_HERSHEY_PLAIN,
                        1.0, (0, 0, 0), thickness=1)
            # cv2.arrowedLine(img, begin_xy, (x, y), (0, 0, 255), 2, 0, 0, 0.1)  # 画完终点后画箭头
            cv2.imshow("image", img)
            print(x, y)
            num  = 1

    cv2.namedWindow('image', cv2.WINDOW_NORMAL)
    cv2.setMouseCallback("image", on_EVENT_LBUTTONDOWN)

    while True:
        cv2.imshow('image', img)
        key = cv2.waitKey(1)
        if key == ord('q'):  # 在键盘上按Q键退出画图
            break
    if num % 2 == 1:  # 如果num为奇数说明有一个起点多余了 去掉
        a = a[:-1]
        b = b[:-1]
    print(a, b)
    return a, b


# 将坐标点列表a,b 转换为corner_list(坐标点必须为(x,y)形式)
def get_corner_list(a, b):
    corner_list = []
    for i in range(int(len(a) / 2)):
        corner_list.append([a[2 * i], b[2 * i], a[2 * i   1], b[2 * i   1]])
    # print(corner_list)
    return corner_list


# 将在选区外的掩膜去除
# 判断点是否在选择区域内
def in_box(i, j, corner_list):
    # if_inbox = False
    for k in corner_list:
        if i >= k[0] and i <= k[2] and j >= k[1] and j <= k[3]:
            return True
        else:
            continue
    return False


def cut(mask_r, corner_list):
    for i in range(mask_r.shape[0]):
        for j in range(mask_r.shape[1]):
            if mask_r[i, j] == 255 and not in_box(j, i, corner_list):
                mask_r[i, j] = 0
            else:
                continue
    return mask_r


# 主函数
def main(corner_list, img_path, lower, upper, color):
    # 转为hsv颜色模式
    img = cv2.imread(img_path)
    hsv = cv2.cvtColor(img, cv2.COLOR_BGR2HSV)
    mask = cv2.inRange(hsv, lower, upper)
    if color == 1:
        lower_red2 = np.array([0, 60, 60])
        upper_red2 = np.array([10, 255, 255])  # thers is two ranges of red
        mask_2 = cv2.inRange(hsv, lower_red2, upper_red2)
        mask = mask   mask_2
    mask = cut(mask, corner_list)
    gray_image = cv2.imread(img_path, cv2.IMREAD_GRAYSCALE)
    gray = cv2.merge([gray_image, gray_image, gray_image])
    # 将mask于原视频帧进行按位与操作,则会把mask中的白色用真实的图像替换:
    res = cv2.bitwise_and(img, img, mask=mask)
    mask_bg = cv2.bitwise_not(mask)
    gray = cv2.bitwise_and(gray, gray, mask=mask_bg)
    result = res   gray
    cv2.namedWindow('Result', 0)
    cv2.imshow('Result', result)
    cv2.imwrite('result.jpg', result)
    cv2.waitKey(0)


if __name__ == '__main__':
    img_path = '../img/img.jpg'
    lower, upper, color = color_choose()
    img = cv2.imread(img_path)
    num = 0
    a, b = choose_range(img)
    cv2.destroyAllWindows()
    corner_list = get_corner_list(a, b)
    main(corner_list, img_path, lower, upper, color)

0 人点赞