Article Directory
foreword
This article is the first of the Image Segmentation column. Image segmentation is a very important part of the entire image processing process. Most of it is used as a preprocessing step of the entire image processing. The purpose of segmentation is mainly to obtain the ROI area and lay a solid foundation for subsequent feature extraction and pattern recognition. .
This article mainly introduces how to use Kmeans method for image segmentation.
1. What is Kmeans?
Kmeans is a simple clustering algorithm that divides the input data into k categories. This algorithm can continuously extract the center point of the current classification, and finally complete the clustering when the classification is stable. Essentially, Kmeans is an iterative algorithm.
The general purpose of the Kmeans algorithm is to classify the data, let the computer analyze the data, and automatically divide the data into k categories. The picture is actually a two-dimensional or three-dimensional array, and the Kmeans classification of the picture will naturally complete the segmentation.
2. How to use Kmeans
1. Kmeans function in opencv
This picture is the parameter analysis of the Kmeans function in opencv-python. There are a lot of information on the Internet, so I won't go into details here.
2. Kmeans code example
Classify one-dimensional data
from sys import flags
from cv2 import cv2
from matplotlib.image import imread
import numpy as np
from matplotlib import pyplot as plt
from numpy.core.defchararray import center
'''
Kmeans对一维数据进行聚类
'''
# 随机生成小米和大米两组数据，小米数值特征范围0到50，大米特征范围200到250，都是60个
xiaomi = np.random.randint(0,50,60)
dami = np.random.randint(200,250,60)
# 将大米小米数据按水平线方向合并，得到（120，）
mi = np.hstack((xiaomi,dami))
# print(mi.dtype)
# print(mi.shape)
# 将mi的形状变成一维向量，并且数据类型变成Kmeans需要的float32
mi = mi.reshape((120,1))
mi = np.float32(mi)
# 设置Kmeans的终止条件
critera = (cv2.TermCriteria_EPS+cv2.TermCriteria_MAX_ITER,10,1.0)
# 设置Kmeans的初始化方式
flags = cv2.KMEANS_RANDOM_CENTERS
# 将数据传入，进行聚类
r,best,center = cv2.kmeans(mi,2,None,criteria=critera,attempts=10,flags=flags)
print(r)
print(best)
print(center)
# 标签为0的是小米，将该结果数据利用逻辑索引取出赋给新变量xiaomi
xm = mi[best==0]
dm = mi[best==1]
# 展示结果
plt.plot(xm,'ro')
plt.plot(dm,'bo')
plt.plot(center[0],'rx')
plt.plot(center[1],'bx')
plt.show()
The result is shown in the figure below
Classify the two-dimensional data
# 每组数据是二维,xiaomi数据是（30,2）的size，且数据范围是0到20，dami的size也是（30,2），范围是40到60
xiaomi = np.random.randint(0,20,(30,2))
dami = np.random.randint(40,60,(30,2))
# 将两种数据按照垂直方式合并，得到（60,2）
mi = np.vstack((xiaomi,dami))
print(mi.dtype)
print(mi.shape)
# 转换数据格式
mi = np.float32(mi)
# 设置Kmeans函数参数
critera = (cv2.TermCriteria_EPS+cv2.TermCriteria_MAX_ITER,10,1.0)
flags = cv2.KMEANS_RANDOM_CENTERS
r,best,center = cv2.kmeans(mi,2,None,criteria=critera,attempts=10,flags=flags)
print(r)
print(best.shape)
print(center)
# 因为是二维数据，得到的标签不是（120，），而是（120,1），是一个一维向量，所以逻辑索引和上面有的区别
xm = mi[best[:,0]==0]
dm = mi[best[:,0]==1]
plt.scatter(xm[:,0],xm[:,1],c='g',marker='s')
plt.scatter(dm[:,0],dm[:,1],c='r',marker='s')
plt.xlabel('height'),plt.ylabel('width')
plt.show()
Segment the picture
# 对图像用kmeans聚类
# 显示图片的函数
def show(winname,src):
cv2.namedWindow(winname,cv2.WINDOW_GUI_NORMAL)
cv2.imshow(winname,src)
cv2.waitKey()
img = cv2.imread('segmentation/1.jpg')
o = img.copy()
print(img.shape)
# 将一个像素点的rgb值作为一个单元处理，这一点很重要
data = img.reshape((-1,3))
print(data.shape)
# 转换数据类型
data = np.float32(data)
# 设置Kmeans参数
critera = (cv2.TermCriteria_EPS+cv2.TermCriteria_MAX_ITER,10,0.1)
flags = cv2.KMEANS_RANDOM_CENTERS
# 对图片进行四分类
r,best,center = cv2.kmeans(data,4,None,criteria=critera,attempts=10,flags=flags)
print(r)
print(best.shape)
print(center)
center = np.uint8(center)
# 将不同分类的数据重新赋予另外一种颜色，实现分割图片
data[best.ravel()==1] = (0,0,0)
data[best.ravel()==0] = (255,0,0)
data[best.ravel()==2] = (0,0,255)
data[best.ravel()==3] = (0,255,0)
# 将结果转换为图片需要的格式
data = np.uint8(data)
oi = data.reshape((img.shape))
# 显示图片
show('img',img)
show('res',oi)
Original image
Result image
Summarize
Kmeans' image segmentation is essentially based on grayscale segmentation. Since it is based on grayscale segmentation, it will be sensitive to illumination changes. But this kind of segmentation is different from the traditional threshold segmentation in that it can directly segment the color image, and it has more adaptability than the traditional threshold segmentation. Of course, the strength of this adaptability and the Otsu algorithm (ostu) remains to be verified.