利用OpenCV对python中岛结构图像进行边缘检测

Question

我在python中遇到了一些图像识别问题。我想在下图中找到单独岛屿的区域：

https://drive.google.com/file/d/1GW6OCTMLtw9d8Opgtq3y4C5xshLP1siz/view?usp=sharing

为了分别找到所有岛屿的面积，我试图找到岛屿的轮廓，然后我计算出面积。我根据轮廓区域的大小为每个轮廓赋予不同的颜色。然而，岛屿的轮廓往往重叠，我无法正确分离它们。在这里，您可以找到不同步骤的图像以及对图像的影响

请参阅：单独的过滤步骤：

我使用的代码（包括注释）如下：

    # -*- coding: utf-8 -*-
"""
Created on Fri Jun 15 12:15:17 2018

@author: Gdehaan
"""


import matplotlib.pyplot as plt
import numpy as np
import glob
import cv2 as cv
from scipy.ndimage.morphology import binary_closing
from scipy.ndimage.morphology import binary_fill_holes




plt.close('all')

#Create a list of the basic colors to draw the contours
all_colors = [(255, 0 , 0), (0, 255 , 0), (0, 0, 255), (255, 0 , 255), (255, 255 , 0), (0, 255 , 255), (0, 0, 0)]


#Here we add random rgb colors to draw the contours later since we might have a lot of contours
col_count = 100
counter = 0
while counter < col_count:
    all_colors.append(tuple(np.random.choice(range(256), size=3)))
    counter+=1

pltcolors = [] #Here we convert the rgb colors to the matplotlib syntax between 0 and 1 instead of between 0 and 255
for i in range(len(all_colors)):
    pltcolors.append(tuple([float(color)/255 for color in all_colors[i]]))

figures = glob.glob('*.tif')



figure_path = 'C:\Users\gdehaan\Desktop\SEM analysis test\zoomed test\{}'

for figure in figures:

    if figure == '80nm.tif':
        fig_title = str(figure.strip('.tif')) #Create a figure title based on the filename
        fig_title_num = int(figure.strip('nm.tif')) #Get the numerical value of the filename (80)
        pixel_scale = 16.5e-3 #Scalefactor for pixel size

    path = figure_path.format(figure)
    img_full = cv.imread(path , 0) #Import figure, 0 = GrayScale
    img = img_full[:880, :1000] #Remove labels etc.
    img_copy = np.copy(img) #create a copy of the image (not needed)


    #Here we create a blanco canvas to draw the contours on later, with the same size as the orignal image
    blanco = np.zeros([int(np.shape(img)[0]), int(np.shape(img)[1]), 3], dtype=np.uint8) 
    blanco.fill(255)

    #We use a bilateral filter to smooth the image while maintaining sharp borders    
    blur = cv.bilateralFilter(img, 6, 75, 75) 

    #Threshold the image to a binary image with a threshold value determined by the average of the surrounding pixels
    thresh = cv.adaptiveThreshold(blur, 255, cv.ADAPTIVE_THRESH_MEAN_C, cv.THRESH_BINARY, 11, 2)

    #Here we fill the holes in the Islands
    hole_structure = np.ones((3,3))
    no_holes= np.array(binary_fill_holes(thresh, structure = hole_structure).astype(int), dtype = 'uint8')

    #Here we close some of the smaller holes still present
    closed = np.array(binary_closing(no_holes).astype(int), dtype = 'uint8')

    #Here we find the contours based on a predetermined algorithm
    im2, contours, hierarchy = cv.findContours(closed, cv.RETR_TREE, cv.CHAIN_APPROX_NONE)

    #Here we calculate the area of all the contours 
    areas = []
    for i in range(len(contours)):
        areas.append(cv.contourArea(contours[i]))
    avg_area = np.mean(areas)

    #Here we sort the contours based on the area they have
    areas_sorted, contours_sorted_tup = zip(*sorted(zip(areas, contours), key = lambda x: x[0]))
    contours_sorted = list(contours_sorted_tup)

    #Here we filter the islands below the average Island size
    contours_sf = []
    areas_sf = []
    for i in range(len(contours_sorted)):
        if areas_sorted[i] > 2*avg_area:
            contours_sf.append(contours_sorted[i])
            areas_sf.append(np.asarray(areas_sorted[i])*(pixel_scale**2))


    #Create the histogram data
    max_bin = max(areas_sf)+3 #Value for the maximal number of bins for the histogram
    num_bins = float(max_bin)/30 #Value for number of bins
    hist_data, bins = np.histogram(areas_sf, np.arange(0, max_bin, num_bins))

    #Create a list of colors matching the bin sizes
    colors_temp = []
    for i,j in enumerate(hist_data):
        colors_temp.append(int(j)*[all_colors[i]])

    #Concatenate the list manually, numpy commands don't work well on list of tuples
    colors = []
    for i in range(len(colors_temp)):
        for j in range(len(colors_temp[i])):
            if colors_temp[i][j] != 0:
                colors.append(colors_temp[i][j])
            else:
                colors.append((0, 0, 0))

    #Here we draw the contours over the blanco canvas       
    for i in range(len(contours_sf)):
        cv.drawContours(blanco, contours_sf[i], -1, colors[i], 2) 

    #The rest of the script is just plotting    
    plt.figure()
    plt.suptitle(fig_title)

    plt.subplot(231)
    plt.title('Raw image')
    plt.imshow(img, 'gray')
    plt.xticks([])
    plt.yticks([])

    plt.subplot(232)
    plt.title('Bilateral filtered')
    plt.imshow(blur, 'gray')
    plt.xticks([])
    plt.yticks([])

    plt.subplot(233)
    plt.title('Thresholded')
    plt.imshow(thresh, 'gray')
    plt.xticks([])
    plt.yticks([])

    plt.subplot(234)
    plt.title('Edges closed & Holes filled')
    plt.imshow(closed, 'gray')
    plt.xticks([])
    plt.yticks([])

    plt.subplot(235)
    plt.title('Contours')
    plt.imshow(blanco)
    plt.xticks([])
    plt.yticks([])

    plt.subplot(236)
    plt.title('Histogram')
    for i in range(len(hist_data)):
        plt.bar(bins[i], hist_data[i], width = bins[1], color = pltcolors[i])
    plt.xlabel(r'Island size ($\mu$m$^{2}$)')
    plt.ylabel('Frequency')
    plt.axvline(x=np.mean(areas_sf), color = 'k', linestyle = '--', linewidth = 3)

    figManager = plt.get_current_fig_manager()
    figManager.window.showMaximized()


    plt.figure()
    plt.suptitle(fig_title, fontsize = 30)

    plt.subplot(121)
    plt.title('Contours' + '\n', linespacing=0.3, fontsize = 20)
    plt.imshow(blanco)
    plt.imshow(img, 'gray', alpha = 0.7)
    plt.xticks([])
    plt.yticks([])

    plt.subplot(122)
    plt.title('Histogram' + '\n', linespacing=0.3, fontsize = 20)
    for i in range(len(hist_data)):
        plt.bar(bins[i], hist_data[i], width = bins[1], color = pltcolors[i])
    plt.xlabel(r'Island size ($\mu$m$^{2}$)', fontsize = 16)
    plt.ylabel('Frequency', fontsize = 16)
    plt.axvline(x=np.mean(areas_sf), color = 'k', linestyle = '--', linewidth = 3)

    figManager = plt.get_current_fig_manager()
    figManager.window.showMaximized()

问题出现在“阈值化”图像到“边缘闭合和孔填充”图像上。从这里看起来很多边缘都熔化在一起。我不能让它们很好地分开，因此我的轮廓开始重叠或根本不被识别。我可以依靠使用一些帮助来更好/有效地分离岛屿。我尝试使用过滤器值，但我没有得到更好的结果。

Answer 1

我尝试了一种稍微不同的方法。看看下面的代码。

注意：用于模糊和形态操作的每个滤镜的内核大小是您可以调整以获得更好结果的参数。我的方法是为了给你指路。此外，我建议使用cv2.imshow()可视化每一步，以便更好地了解正在发生的事情。

码：

im = cv2.imread('80nm.tif')
imgray = cv2.cvtColor(im,cv2.COLOR_BGR2GRAY)

#--- Bilateral filtering ---
blur = cv2.bilateralFilter(imgray, 6, 15, 15)

#--- Perform Otsu threshold ---
ret, otsu_th = cv2.threshold(blur, 0, 255, cv2.THRESH_BINARY | cv2.THRESH_OTSU)   

接下来我使用了Watershed implementation of OpenCV的一些步骤

#--- noise removal ---
kernel = np.ones((3, 3), np.uint8)
opening = cv2.morphologyEx(otsu_th, cv2.MORPH_OPEN, cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (3, 3)), iterations = 2)

#--- sure background area ---
sure_bg = cv2.dilate(opening, kernel, iterations = 1)
cv2.imshow('sure_bg', sure_bg)

#--- Finding sure foreground area ---
dist_transform = cv2.distanceTransform(opening, cv2.DIST_L2, 5)
ret, sure_fg = cv2.threshold(dist_transform, 0.1 * dist_transform.max(), 255, 0)

cv2.normalize(dist_transform, dist_transform, 0, 1, cv2.NORM_MINMAX, dtype=cv2.CV_32F)
#cv2.imshow('dist_transform_normalized', dist_transform)
#cv2.imshow('sure_fg', sure_fg)

#--- Finding unknown region ---
sure_fg = np.uint8(sure_fg)
unknown = cv2.subtract(opening, sure_fg)
cv2.imshow('unknown', unknown)