主要思路就是:
创新互联基于成都重庆香港及美国等地区分布式IDC机房数据中心构建的电信大带宽,联通大带宽,移动大带宽,多线BGP大带宽租用,是为众多客户提供专业成都电信服务器托管报价,主机托管价格性价比高,为金融证券行业服务器托管,ai人工智能服务器托管提供bgp线路100M独享,G口带宽及机柜租用的专业成都idc公司。import cv2
import numpy as np
import matplotlib.pyplot as plt
import random
import plotly.express as px
from queue import Queue,PriorityQueue
import sys
class WATERSHED:
def __init__(self,img,markers):
if img.ndim!=2:
raise ValueError("请输入灰度图")
self.img = img
# 求图像梯度幅值
self.gradient()
self.markers = markers.astype(np.int32)
self.img_highest = np.max(self.img)
self.img_lowest = np.min(self.img)
self.WSHED = -1# 分水岭
self.IN_QUEUE = -2#已入队,但未赋予标记
# 根据梯度幅值图的灰度大最小值创建有序队列
self.ordered_queue = [[] for _ in range(self.img_highest-self.img_lowest+1)]
def gradient(self):
# self.img = cv2.Canny(self.img,80,150)
d_x= cv2.Sobel(self.img,cv2.CV_64F,1,0,ksize=3)
d_y= cv2.Sobel(self.img,cv2.CV_64F,0,1,ksize=3)
value = cv2.magnitude(d_x,d_y)
self.img = cv2.convertScaleAbs(value)
def ordered_queue_push(self,i,j):
gray_leve = self.img[i][j]
self.ordered_queue[gray_leve-self.img_lowest].append([i,j])
self.markers[i][j] = self.IN_QUEUE
def PriorityQueue_push(self,water_stage):
# 将等于水位高度的像素点全部入优先队列
for pix_coord in self.ordered_queue[water_stage-self.img_lowest]:
self.wait_water_queue.put((water_stage,pix_coord))
def wait_water_queue_push(self,x,y):
self.wait_water_queue.put((self.img[x][y],[x,y]))
self.markers[x][y] = self.IN_QUEUE
def get_label(self,pix_coord:list,water_stage):
x,y = pix_coord
neighbour_pix_values = {}
for i in [-1,0,1]:
for j in [-1,0,1]:
pix_x,pix_y = x+i,y+j
if pix_x<0 or pix_y<0 or pix_x >=self.img.shape[0] \
or pix_y >=self.img.shape[1] :
continue
# 必须以四邻域搜索,八邻域会越过边界
if abs(i+j)==1:
if neighbour_pix_values.get(self.markers[pix_x][pix_y])!=None:
neighbour_pix_values[self.markers[pix_x][pix_y]]+=1
else:
neighbour_pix_values[self.markers[pix_x][pix_y]]=1
if self.markers[pix_x][pix_y]==0:#不在队列且无标签
if self.img[pix_x][pix_y]<=water_stage:
#低于水位,入优先队列
self.wait_water_queue_push(pix_x,pix_y)
else:
#高于水位,还淹不到
self.ordered_queue_push(pix_x,pix_y)
pix_values = list(neighbour_pix_values)
#去掉分水岭、入队标记和未标记区域值
pix_values = [value for value in pix_values \
if value>0 and value!=self.IN_QUEUE]
if len(pix_values)==1:
self.markers[x][y]=pix_values[0]
else :
self.markers[x][y]=self.WSHED
def run(self):
for i in range(self.img.shape[0]):
for j in range(self.img.shape[1]):
#已标记区域全部入队
if self.markers[i][j] >0:
gray_leve = self.img[i][j]
self.ordered_queue[gray_leve-self.img_lowest].append([i,j])
for water_stage in range(self.img_lowest,self.img_highest+1):
sys.stdout.write("水位高度{}".format(water_stage))
sys.stdout.flush()
self.wait_water_queue = PriorityQueue()
#将水位为water_stage的全部入优先队列
self.PriorityQueue_push(water_stage)
#每轮循环结束小于等于water_stage的像素值都得清空
while self.wait_water_queue.qsize() >0:
pix_coord = self.wait_water_queue.get()[1]
self.get_label(pix_coord,water_stage)
# 动画
plt.clf()
plt.imshow(self.markers)
plt.pause(0.001)
return self.markers
2 获取初始标记如果如opencv官方例程那样,用距离变换来确认标记,非常容易过分割,所以我们可以通过手动进行标记。
可以使用opencv的鼠标事件实现标记。
import cv2
import numpy as np
class draw_markers:
def __init__(self,img,resize_rate:int=1):
if img.ndim == 2:
self.img = cv2.cvtColor(img,cv2.COLOR_GRAY2BGR)
else:
self.img = img
self. resize_rate = resize_rate
self.r,self.w = self.img.shape[:2]
self.markers = np.zeros_like(self.img)
# 放大
self.img = cv2.resize(self.img,(self.r*resize_rate,self.w*resize_rate))
self.markers = cv2.resize(self.markers,(self.r*resize_rate,self.w*resize_rate))
self.cls_num =0
#标记颜色列表
self.color_list = [(255,0,0),(0,255,0),(0,0,255),(255,255,0),(255,0,255),(0,255,255)]
self.drawing = False
def mark(self,event, x, y,flags, param):
if event==cv2.EVENT_LBUTTONDOWN:
self.drawing = True
if event == cv2.EVENT_LBUTTONUP:
self.drawing = False
if self.drawing:
cv2.circle(self.img,(x,y),radius=1,color=self.color_list[self.cls_num],thickness=-1)
cv2.circle(self.markers,(x,y),radius=1,color=self.color_list[self.cls_num],thickness=-1)
def draw(self):
cv2.namedWindow('image')
cv2.setMouseCallback('image',self.mark)
while True:
cv2.imshow('image',self.img)
k = cv2.waitKey(1) & 0xFF
if k==ord('c'):
self.cls_num +=1
print("第{}类".format(self.cls_num+1))
if k==27:
break
cv2.destroyAllWindows()
result = cv2.cvtColor(self.markers,cv2.COLOR_BGR2GRAY)
#不能用插值,会改变标签值
result = result[::self.resize_rate,::self.resize_rate]
cv2.imwrite('./markers.png',self.img)
return result
3 主函数要想分割的好,要避免不同区域的连通,相同区域的分裂。锐化和高斯模糊感觉用了效果不好。
img = cv2.imread('test.png',0)
#标记图像
DM = draw_markers(img,1)
markers = DM.draw()
#开运算
kernel = np.ones((5,5),dtype=np.uint8)
img = cv2.morphologyEx(img,cv2.MORPH_OPEN,kernel,iterations=1)
# img = cv2.erode(img,kernel=np.ones((5,5)),iterations=2)
# img = cv2.GaussianBlur(img,(5,5),sigmaX=0)
# #锐化
# kernel = np.array([[1, 1, 1], [1, -8, 1], [1, 1, 1]], dtype=np.float32)
# imgLaplacian = cv2.filter2D(img, cv2.CV_32F, kernel)
# sharp =np.float32(img)
# imgResult = sharp - imgLaplacian
# imgResult = np.clip(imgResult,0,255)
# imgResult = np.uint8(imgResult)
#开始分水岭算法
WSHED= watershed.WATERSHED(img,markers)
result = WSHED.run()
#将所有的分水岭转为255
result[result==-1]==255
result = result.astype(np.uint8)
imshow(result,'result.png')
4 实验结果
4.1 标记图像4.2 分割结果对比opencv官方例程,过分割现象减少。
opencv例程结果如下:
COLOR IMAGE SEGMENTATION
Image Segmentation with Distance Transform and Watershed Algorithm
OpenCV cv::watershed 分水岭算法论文解读以及numpy实现
你是否还在寻找稳定的海外服务器提供商?创新互联www.cdcxhl.cn海外机房具备T级流量清洗系统配攻击溯源,准确流量调度确保服务器高可用性,企业级服务器适合批量采购,新人活动首月15元起,快前往官网查看详情吧