图像处理-007形态变换(二)

Hit-and-Miss Transform(击中击不中)

Hit-and-Miss从二值图像中查找前景色/背景色的特定模式非常有用，与腐蚀，膨胀使用的kernel仅存在0，1不同，hit-and-Misss的kernel中除0,1外还存在空白。腐蚀，膨胀，开操作，闭操作，形态学梯度，顶帽，黑帽运算使用的kernel中1用来表示前景色，0表示不关注，在hit-and-miss运算中的1表示前景色，0表示背景色，空白表示不关注。

以图006-14中展示的kernel为例，该kernel可用来找到二值图像中的角点，运算时将kernel置于图像之上，锚点依次与图像中的坐标P对应，P的邻域元素依次与kernel的邻域元素比较，除去kernel中的空白元素外，其余元素完全匹配时则P标记为1，否则为0.

图007-1 查找图像角点的kernel

使用图007-2中的kernel找出图像中的四个角点如图007-3所示

图007-2 寻找图像四角点的kernel

图007-3 二值图像Hit-And-Miss运算结果图

实现代码: C++

/*** hit and miss* @param img* @return*/
int Morphological::hit_and_miss_transform(const Mat &img) {logger_info("======hit_and_miss_transform=====================");//  kernel size为7*7Mat kernel = getStructuringElement(MORPH_ELLIPSE, Size(7, 7));Mat dst;morphologyEx(img, dst, MORPH_HITMISS, kernel);imshow("hit_and_miss", dst);return EXIT_SUCCESS;
}

实现代码:python

# hit and miss
def hit_and_miss_transform(origin_img):# 非灰度图像转化维灰度图像if origin_img.shape[2] == 3:gray_image = cv.cvtColor(origin_img, cv.COLOR_BGR2GRAY)else:gray_image = origin_imgtitles = ["origin_image"]images = [cv.cvtColor(gray_image, cv.COLOR_BGR2RGB)]for i in range(3, 13, 2):kernel = cv.getStructuringElement(cv.MORPH_RECT, (i, i))dst = cv.morphologyEx(gray_image, cv.MORPH_HITMISS, kernel)titles.append("hit_and_miss_kernel_" + str(i))images.append(cv.cvtColor(dst, cv.COLOR_BGR2RGB))for i in range(6):plt.subplot(2, 3, i + 1)plt.imshow(images[i], 'gray', vmin=0, vmax=255)plt.title(titles[i])plt.xticks([]), plt.yticks([])plt.show()

图007-4 二值图像不同kernel大小下hit-and-miss效果图

Thining(细化)

细化会删除二值图像中选定的前景(白色)像素，类似于腐蚀、开操作，删除不必要的像素点，仅保留图像的轮廓。

dst=thining(src,element)=src−hit-and-miss(src,element)\texttt{dst} = \mathrm{thining} ( \texttt{src} , \texttt{element} )= \texttt{src} - \texttt{hit-and-miss} ( \texttt{src} , \texttt{element} ) dst=thining(src,element)=src−hit-and-miss(src,element)

图像细化后依然保持图像的骨架且细小部分的连通性不变，细化将图像线条从多像素宽度减少至单位像素宽度。

opencv提供函数thinning() 用于图像的细化。其详细描述如下:

void cv::ximgproc::thinning(InputArray     src,    #输入图像 8bit单通道灰度图像OutputArray     dst, #输出图像 与输入图像同类型，同大小int     thinningType = THINNING_ZHANGSUEN #细化类型)
Python:
cv.ximgproc.thinning(src[, dst[, thinningType]]    )->dst

实现代码: C++

/*** 图像细化 THINNING_ZHANGSUEN* @param img* @return*/
int Morphological::thinning_ZHANGSUEN_transform(const Mat &img) {logger_info("======hit_and_miss_transform=====================");Mat dst;ximgproc::thinning(img, dst, ximgproc::ThinningTypes::THINNING_ZHANGSUEN);imshow("thinning zhangsuen", dst);return EXIT_SUCCESS;
}/*** 图像细化 THINNING_GUOHALL* @param img* @return*/
int Morphological::thinning_GUOHALL_transform(const Mat &img) {logger_info("======thinning_GUOHALL_transform=====================");Mat dst;ximgproc::thinning(img, dst, ximgproc::ThinningTypes::THINNING_GUOHALL);imshow("thinning guohall", dst);return EXIT_SUCCESS;
}

实现代码: python

# thinning
def thinning_transform(origin_img):# 非灰度图像转化维灰度图像if origin_img.shape[2] == 3:gray_image = cv.cvtColor(origin_img, cv.COLOR_BGR2GRAY)else:gray_image = origin_imgtitles = ["origin_image"]# images = [cv.cvtColor(gray_image, cv.COLOR_BGR2RGB)]images = [gray_image]dst = cv.ximgproc.thinning(gray_image, cv.ximgproc.THINNING_ZHANGSUEN)images.append(dst)titles.append("thinning_zhangsuen")dst = cv.ximgproc.thinning(gray_image, cv.ximgproc.THINNING_GUOHALL)images.append(dst)titles.append("thinning_gauhall")for i in range(3):plt.subplot(1, 3, i + 1)plt.imshow(images[i], 'gray', vmin=0, vmax=255)plt.title(titles[i])plt.xticks([]), plt.yticks([])plt.show()

细化效果图如图007-5所示

图007-5 图像细化效果图

Thickening(粗化)

粗化是形态学的相对操作，细化前景色相当于粗化背景色。粗化操作由源图像与hit-and-miss处理后的并集组成。

dst=thining(src,element)=src∪hit-and-miss(src,element)\texttt{dst} = \mathrm{thining} ( \texttt{src} , \texttt{element} )= \texttt{src} \cup \texttt{hit-and-miss} ( \texttt{src} , \texttt{element} ) dst=thining(src,element)=src∪hit-and-miss(src,element)

Skeletonization/Medial Axis Transform

骨架化指将二值图像的前景色(高亮部分)减少直至骨架的过程，该过程丢弃原始图像中大量的前景像素，但保留了图像的原始区域范围与连通性。

骨架化后的图像既然保留了原始图像的骨架及连通性，则在视觉处理中可大幅减少数据量的计算和对内存的占用。

实现代码: C++

/*** 骨架* @param img* @return*/
int Morphological::skeleton_transform(const Mat &img) {logger_info("======skeleton_transform=====================");imshow("image", img);Mat dst;
//  图像阈值转化  灰度图像转二值图像threshold(img, dst, 127, 255, THRESH_BINARY);
// 卷积核Mat kernel = getStructuringElement(MORPH_CROSS, Size(7, 7));Mat skel(dst.size(), CV_8UC1, cv::Scalar(0));Mat temp(dst.size(), CV_8UC1);bool done;do {cv::morphologyEx(dst, temp, MORPH_OPEN, kernel);cv::bitwise_not(temp, temp);cv::bitwise_and(dst, temp, temp);cv::bitwise_or(skel, temp, skel);cv::erode(dst, dst, kernel);double max;cv::minMaxLoc(dst, 0, &max);done = (max == 0);} while (!done);imshow("skeleton", skel);return EXIT_SUCCESS;
}

实现代码:python


# 骨架化
def skeleton_transform(origin_img):# 非灰度图像转化维灰度图像if origin_img.shape[2] == 3:gray_image = cv.cvtColor(origin_img, cv.COLOR_BGR2GRAY)else:gray_image = origin_imgtitles = ["origin_image"]images = [gray_image]ret, binary_img = cv.threshold(gray_image, 127, 255, cv.THRESH_BINARY)titles.append("binary_image")images.append(binary_img)# 获取卷积核kernel = cv.getStructuringElement(cv.MORPH_CROSS, (7, 7))skeleton_img = np.zeros(binary_img.shape, np.uint8)# tmp_img = np.zeros(binary_img.shape, np.uint8)while True:# 开运算tmp_img = cv.morphologyEx(binary_img, cv.MORPH_OPEN, kernel)tmp_img = cv.bitwise_not(tmp_img)tmp_img = cv.bitwise_and(binary_img, tmp_img)skeleton_img = cv.bitwise_or(skeleton_img, binary_img)binary_img = cv.erode(binary_img, kernel)min, max, _, _ = cv.minMaxLoc(binary_img)if max == 0:breaktitles.append("skeleton_image")images.append(skeleton_img)for i in range(3):plt.subplot(1, 3, i + 1)plt.imshow(images[i], 'gray', vmin=0, vmax=255)plt.title(titles[i])plt.xticks(), plt.yticks([])plt.show()

参考文献：

https://homepages.inf.ed.ac.uk/rbf/HIPR2/morops.htm
http://www.cs.cmu.edu/~cil/vision.html
https://docs.opencv.org/4.6.0/db/d06/tutorial_hitOrMiss.html
https://felix.abecassis.me/2011/09/opencv-morphological-skeleton/