实战 | 粘连物体分割与计数应用(二)--基于距离变换+分水岭算法 Halcon/OpenCV实现比较

导读

本文主要介绍基于基于距离变换分水岭算法实现粘连物体的分割与计数方法，并对比Halcon与OpenCV实现差异。

背景介绍

在实际的视觉应用场景中，我们常常会遇到物体/元件的计数问题，而计数时比较常见的情形就是物体相邻或粘连，对相邻或粘连物体的分割将直接影响着最终计数的准确性。后面将分篇介绍粘连物体分割计数的常用方法，包括：

【1】形态学连通域处理方法

【2】距离变换分水岭分割方法

【3】其他方法

本文将对第【2】种方法分别用Halcon和OpenCV实现并做简单对比。

实例演示与实现步骤

* 实例一：糖豆分割与计数

测试图像(图片来源--Halcon例程图)：

实现步骤：

【1】阈值处理：区间固定阈值或OTSU阈值

【2】距离变换

【3】分水岭算法

【4】求区域重叠部分(交集)

【5】结果标记与显示

Halcon实现代码：

代码语言：javascript复制

dev_clear_window()
dev_get_window (WindowHandle)
*读取图片
read_image (Image, 'pellets')
get_image_size (Image, Width, Height)
*自动阈值分割
rgb1_to_gray (Image, GrayImage)
threshold (GrayImage, Regions, 110, 255)
connection (Regions, ConnectedRegions)
select_shape (ConnectedRegions, SelectedRegions, 'area', 'and', 40, 50000)

*欧式距离函数的距离变换
distance_transform (SelectedRegions, DistanceImage, 'octagonal', 'true', Width, Height)
*int4转byte
convert_image_type (DistanceImage, ImageConverted, 'byte')
*图像取反
invert_image (ImageConverted, ImageInvert)

*图像比例增强 按最大比例增强对比度
scale_image_max (ImageInvert, ImageScaleMax)

*分水岭算法
watersheds_threshold (ImageScaleMax, Basins, 30)
select_shape (Basins, SelectedBasins, 'area', 'and', 2000, 50000)
gen_contour_region_xld (SelectedBasins, Contours, 'border')

*取出两个区域中重叠的部分
intersection (SelectedBasins, SelectedRegions, RegionIntersection)

dev_display(Image)
dev_display(RegionIntersection)
dev_display(Image)
dev_set_draw ('margin')
dev_set_line_width (3)
dev_display (RegionIntersection)
dev_set_color ('red')
dev_set_line_width (2)
area_center (RegionIntersection, Area, Row, Column)
gen_cross_contour_xld (Cross, Row, Column, 15, 0.785398)
count_obj (RegionIntersection, Number)
*设置字体颜色
dev_set_color ('green')
*设置文字大小
set_display_font (WindowHandle, 30, 'mono', 'true', 'false')
*设置文字位置
set_tposition (WindowHandle, 15, 220)
write_string(WindowHandle, 'count='   Number)

OpenCV实现代码与效果：

代码语言：javascript复制

#公众号：OpenCV与AI深度学习
#作者：Color Space
import cv2
import numpy as np
import random as rd

def watershed_algorithm(image):
    src = image.copy()
    # 边缘保留滤波EPF  去噪
    blur = cv2.pyrMeanShiftFiltering(image,sp=21,sr=55)
    cv2.imshow("blur", blur)
    # 转成灰度图像
    gray = cv2.cvtColor(blur, cv2.COLOR_BGR2GRAY)
    # 得到二值图像区间阈值
    ret, binary = cv2.threshold(gray, 0, 255, cv2.THRESH_BINARY   cv2.THRESH_OTSU)
    cv2.imshow('thres image', binary)

    # 距离变换
    dist = cv2.distanceTransform(binary, cv2.DIST_L2, 3)
    dist_out = cv2.normalize(dist, 0, 1.0, cv2.NORM_MINMAX)
    cv2.imshow('distance-Transform', dist_out * 100)
    ret, surface = cv2.threshold(dist_out, 0.5*dist_out.max(), 255, cv2.THRESH_BINARY)
    cv2.imshow('surface', surface)
    sure_fg = np.uint8(surface)# 转成8位整型
    cv2.imshow('Sure foreground', sure_fg)

    # Marker labelling
    ret, markers = cv2.connectedComponents(sure_fg)  # 连通区域
    print(ret)
    markers = markers   1 #整个图 1，使背景不是0而是1值

    # 未知区域标记(不能确定是前景还是背景)
    kernel = np.ones((3, 3), np.uint8)
    binary = cv2.morphologyEx(binary, cv2.MORPH_DILATE, kernel, iterations=1)
    unknown = binary - sure_fg
    cv2.imshow('unknown',unknown)

    # 未知区域标记为0
    markers[unknown == 255] = 0
    # 区域标记结果
    markers_show = np.uint8(markers)
    cv2.imshow('markers',markers_show*100)

    # 分水岭算法分割
    markers = cv2.watershed(image, markers=markers)
    min_val, max_val, min_loc, max_loc = cv2.minMaxLoc(markers)
    markers_8u = np.uint8(markers)
    #print(max_val)
    colors = [(255,0,0), (0,255,0), (0,0,255), (255,255,0),
              (255,0,255), (0,255,255), (255,128,0), (255,0,128),
              (128,255,0), (128,0,255), (255,128,128), (128,255,255)]
    for i in range(2,int(max_val 1)):
        ret, thres1 = cv2.threshold(markers_8u, i-1, 255, cv2.THRESH_BINARY)
        ret2, thres2 = cv2.threshold(markers_8u, i, 255, cv2.THRESH_BINARY)
        mask = thres1 - thres2
        cv2.imshow('mask',mask)
        #color = (rd.randint(0,255), rd.randint(0,255), rd.randint(0,255))
        #image[markers == i] = [rd.randint(0,255), rd.randint(0,255), rd.randint(0,255)]
        #image[markers == i] = [colors[i-2]]
        
        contours,hierarchy = cv2.findContours(mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_NONE)
        cv2.drawContours(image,contours,-1,colors[(i-2)],-1)
        #cv2.drawContours(src,contours,-1,colors[(i-2)],-1)
        M = cv2.moments(contours[0])
        cx = int(M['m10']/M['m00'])
        cy = int(M['m01']/M['m00'])#轮廓重心
        cv2.drawMarker(image, (cx,cy),(0,0,255),1,10,2)
        cv2.drawMarker(src, (cx,cy),(0,0,255),1,10,2)
        
    cv2.putText(src,"count=%d"%(int(max_val-1)),(220,30),0,1,(0,255,0),2)
    cv2.putText(image,"count=%d"%(int(max_val-1)),(220,30),0,1,(0,255,0),2)
    cv2.imshow('regions', image)
    result = cv2.addWeighted(src,0.6,image,0.5,0) #图像权重叠加
    cv2.imshow('result', result)

src = cv2.imread('./pellets.png')
cv2.imshow('src', src)
watershed_algorithm(src)
cv2.waitKey(0)
cv2.destroyAllWindows()

OpenCV标记结果：