OpenCV 表盘指针自动读数
前段时间参加了一个表盘指针读数的比赛，今天来总结一下
数据集一共有一千张图片：
方法一：径向灰度求和
基本原理：
将图像以表盘圆心转换成极坐标，然后通过矩阵按行求和找到二值图最大值即为指针尖端
导入需要用到的包
import cv2 as cv
import numpy as np
import math
from matplotlib import pyplot as plt
import os
图像预处理
去除背景：利用提取红色实现
def extract_red(image):
"""
通过红色过滤提取出指针
"""
red_lower1 = np.array([0, 43, 46])
red_upper1 = np.array([10, 255, 255])
red_lower2 = np.array([156, 43, 46])
red_upper2 = np.array([180, 255, 255])
dst = cv.cvtColor(image, cv.COLOR_BGR2HSV)
mask1 = cv.inRange(dst, lowerb=red_lower1, upperb=red_upper1)
mask2 = cv.inRange(dst, lowerb=red_lower2, upperb=red_upper2)
mask = cv.add(mask1, mask2)
return mask
获得钟表中心：轮廓查找，取出轮廓的外接矩形，根据矩形面积找出圆心
def get_center(image):
"""
获取钟表中心
"""
edg_output = cv.Canny(image, 100, 150, 2) # canny算子提取边缘
cv.imshow('dsd', edg_output)
# 获取图片轮廓
contours, hireachy = cv.findContours(edg_output, cv.RETR_TREE, cv.CHAIN_APPROX_SIMPLE)
center = []
cut=[0, 0]
for i, contour in enumerate(contours):
x, y, w, h = cv.boundingRect(contour) # 外接矩形
area = w * h # 面积
if area < 100 or area > 4000:
continue
cv.rectangle(image, (x, y), (x + w, y + h), (255, 0, 0), 1)
cx = w / 2
cy = h / 2
cv.circle(image, (np.int(x + cx), np.int(y + cy)), 1, (255, 0, 0)) ## 在图上标出圆心
center = [np.int(x + cx), np.int(y + cy)]
break
return center[::-1]
由上面的图像可以看出，圆心定位还是非常准确的
图片裁剪
def ChangeImage(image):
"""
图像裁剪
"""
# 指针提取
mask = extract_red(image)
mask = cv.medianBlur(mask,ksize=5)#去噪
# 获取中心
center = get_center(mask)
# 去除多余黑色边框
[y, x] = center
cut = mask[y-300:y+300, x-300:x+300]
# 因为mask处理后已经是二值图像，故不用转化为灰度图像
return cut
剪裁后的图像如下图所示：
极坐标转换
注意：需要将图片裁剪成正方形
def polar(image):
"""
转换成极坐标
"""
x, y = 300, 300
maxRadius = 300*math.sqrt(2)
linear_polar = cv.linearPolar(image, (y, x), maxRadius, cv.WARP_FILL_OUTLIERS + cv.INTER_LINEAR)
mypolar = linear_polar.copy()
#将图片调整为从0度开始
mypolar[:150, :] = linear_polar[450:, :]
mypolar[150:, :] = linear_polar[:450, :]
cv.imshow("linear_polar", linear_polar)
cv.imshow("mypolar", mypolar)
return mypolar
由图像就可以很容易发现指针的顶点
计算角度
def Get_Reading(sumdata):
"""
读数并输出
"""
peak = []
# s记录遍历时波是否在上升
s = sumdata[0] < sumdata[1]
for i in range(599):
# 上升阶段
if s==True and sumdata[i] > sumdata[i+1] and sumdata[i] > 70000:
peak.append(sumdata[i])
s=False
# 下降阶段
if s==False and sumdata[i] < sumdata[i+1]:
s=True
peak.sort()
a = sumdata[0]
b = sumdata[-1]
if not peak or max(a,b) > peak[-1]:
peak.append(max(a,b))
longindex = (sumdata.index(peak[-1]))%599
longnum = (longindex + 1)//25*50
# 先初始化和长的同一刻度
#shortindex = longindex
shortnum = round(longindex / 6)
try:
shortindex = sumdata.index(peak[-2])
shortnum = round(shortindex / 6)
except IndexError:
i=0
while i<300:
i += 1
l = sumdata[(longindex-i)%600]
r = sumdata[(longindex+i)%600]
possibleshort = max(l,r)
# 在短指针可能范围内寻找插值符合条件的值
if possibleshort > 80000:
continue
elif possibleshort < 60000:
break
else:
if abs(l-r) > 17800:
shortindex = sumdata.index(possibleshort) - 1
shortnum = round(shortindex / 6)
break
return [longnum,shortnum%100]
def test():
"""
RGS法测试
"""
image = cv.imread("./BONC/1_{0:0>4d}".format(400) + ".jpg")
newimg = ChangeImage(image)
polarimg = polar(newimg)
psum = polarimg.sum(axis=1, dtype = 'int32')
result = Get_Reading(list(psum))
print(result)
if __name__ == "__main__":
test()
k = cv.waitKey(0)
if k == 27:
cv.destroyAllWindows()
elif k == ord('s'):
cv.imwrite('new.jpg', src)
cv.destroyAllWindows()
[1050, 44]
方法二：Hough直线检测
原理：利用Hough变换检测出指针的两条边，从而两条边的中线角度即为指针刻度
数据预处理与上面的方法类似
可以看到分别检测出了两个指针的左右两条边，然后可以由这四个角度算出两个指针中线的角度，具体计算过程写的有点复杂
class Apparatus:
def __init__(self, name):
self.name = name
self.angle = []
self.src = cv.imread(name)
def line_detect_possible_demo(self, image, center, tg):
'''
:param image: 二值图
:param center: 圆心
:param tg: 直线检测maxLineGap
'''
res = {} # 存放线段的斜率和信息
edges = cv.Canny(image, 50, 150, apertureSize=7)
cv.imshow("abcdefg", edges)
lines = cv.HoughLinesP(edges, 1, np.pi/360, 13, minLineLength=20, maxLineGap=tg)
for line in lines:
x_1, y_1, x_2, y_2 = line[0]
# 将坐标原点移动到圆心
x1 = x_1 - center[0]
y1 = center[1] - y_1
x2 = x_2 - center[0]
y2 = center[1] - y_2
# 计算斜率
if x2 - x1 == 0:
k = float('inf')
else:
k = (y2-y1)/(x2-x1)
d1 = np.sqrt(max(abs(x2), abs(x1)) ** 2 + (max(abs(y2), abs(y1))) ** 2) # 线段长度
d2 = np.sqrt(min(abs(x2), abs(x1)) ** 2 + (min(abs(y2), abs(y1))) ** 2)
# 将长指针与短指针做标记
if d1 < 155 and d1 > 148 and d2 > 115:
res[k] = [1]
elif d1 < 110 and d1 > 100 and d2 > 75:
res[k] = [2]
else:
continue
res[k].append(1) if (x2 + x1) /2 > 0 else res[k].append(0) # 将14象限与23象限分离
cv.line(self.src, (x1 + center[0], center[1] - y1), (x2 + center[0], center[1] - y2), (255, 0, 0), 1)
cv.imshow("line_detect-posssible_demo", self.src)
# 计算线段中点的梯度来判断是指针的左侧线段还是右侧线段
middle_x = int((x_1 + x_2) / 2)
middle_y = int((y_1 + y_2) / 2)
grad_mat = image[middle_y-5:middle_y+6, middle_x-5:middle_x+6]
cv.imshow("grad_mat", grad_mat)
grad_x = cv.Sobel(grad_mat, cv.CV_32F, 1, 0)
grad_y = cv.Sobel(grad_mat, cv.CV_32F, 0, 1)
gradx = np.max(grad_x) if np.max(grad_x) != 0 else np.min(grad_x)
grady = np.max(grad_y) if np.max(grad_y) != 0 else np.min(grad_y)
if ((gradx >=0 and grady >= 0) or (gradx <= 0 and grady >= 0)) and res[k][1] == 1:
res[k].append(1) # 右测
elif ((gradx <= 0 and grady <= 0) or (gradx >= 0 and grady <= 0)) and res[k][1] == 0:
res[k].append(1)
else:
res[k].append(0) # 左侧
# 计算角度
angle1 = [i for i in res if res[i][0] == 1]
angle2 = [i for i in res if res[i][0] == 2]
# 长指针
a = np.arctan(angle1[0])
b = np.arctan(angle1[1])
if a * b < 0 and max(abs(a), abs(b)) > np.pi / 4:
if a + b < 0:
self.angle.append(math.degrees(-(a + b) / 2)) if res[angle1[1]][1] == 1 else self.angle.append(
math.degrees(-(a + b) / 2) + 180)
else:
self.angle.append(math.degrees(np.pi - (a + b) / 2)) if res[angle1[1]][1] == 1 else self.angle.append(
math.degrees(np.pi - (a + b) / 2) + 180)
else:
self.angle.append(math.degrees(np.pi / 2 - (a + b) / 2)) if res[angle1[1]][1] == 1 else self.angle.append(math.degrees(np.pi / 2 - (a + b) / 2) + 180)
print('长指针读数：%f' % self.angle[0])
# 短指针
a = np.arctan(angle2[0])
b = np.arctan(angle2[1])
if a * b < 0 and max(abs(a), abs(b)) > np.pi / 4:
if a + b < 0:
self.angle.append(math.degrees(-(a + b) / 2)) if res[angle2[1]][1] == 1 else self.angle.append(
math.degrees(-(a + b) / 2) + 180)
else:
self.angle.append(math.degrees(np.pi - (a + b) / 2)) if res[angle2[1]][1] == 1 else self.angle.append(
math.degrees(np.pi - (a + b) / 2) + 180)
else:
self.angle.append(math.degrees(np.pi / 2 - (a + b) / 2)) if res[angle2[1]][1] == 1 else self.angle.append(math.degrees(np.pi / 2 - (a + b) / 2) + 180)
print('短指针读数：%f' % self.angle[1])
def get_center(self, mask):
edg_output = cv.Canny(mask, 66, 150, 2)
cv.imshow('edg', edg_output)
# 外接矩形
contours, hireachy = cv.findContours(edg_output, cv.RETR_TREE, cv.CHAIN_APPROX_SIMPLE)
center = []
for i, contour in enumerate(contours):
x, y, w, h = cv.boundingRect(contour) # 外接矩形
area = w * h # 面积
if area > 1000 or area < 40:
continue
#print(area)
# cv.circle(src, (np.int(cx), np.int(cy)), 3, (255), -1)
cv.rectangle(self.src, (x, y), (x + w, y + h), (255, 0, 0), 1)
cx = w / 2
cy = h / 2
cv.circle(self.src, (np.int(x + cx), np.int(y + cy)), 1, (255, 0, 0))
center.extend([np.int(x + cx), np.int(y + cy)])
break
cv.imshow('center', self.src)
return center
def extract(self, image):
red_lower1 = np.array([0, 43, 46])
red_lower2 = np.array([156, 43, 46])
red_upper1 = np.array([10, 255, 255])
red_upper2 = np.array([180, 255, 255])
frame = cv.cvtColor(image, cv.COLOR_BGR2HSV)
mask1 = cv.inRange(frame, lowerb=red_lower1, upperb=red_upper1)
mask2 = cv.inRange(frame, lowerb=red_lower2, upperb=red_upper2)
mask = cv.add(mask1, mask2)
mask = cv.bitwise_not(mask)
cv.imshow('mask', mask)
return mask
def test(self):
self.src = cv.resize(self.src, dsize=None, fx=0.5, fy=0.5) # 此处可以修改插值方式interpolation
mask = self.extract(self.src)
mask = cv.medianBlur(mask, ksize=5) # 去噪
# 获取中心
center = self.get_center(mask)
# 去除多余黑色边框
[y, x] = center
mask = mask[x - 155:x + 155, y - 155:y + 155]
cv.imshow('mask', mask)
#self.find_short(center, mask)
try:
self.line_detect_possible_demo(mask, center, 20)
except IndexError:
try:
self.src = cv.imread(self.name)
self.src = cv.resize(self.src, dsize=None, fx=0.5, fy=0.5) # 此处可以修改插值方式interpolation
self.src = cv.convertScaleAbs(self.src, alpha=1.4, beta=0)
blur = cv.pyrMeanShiftFiltering(self.src, 10, 17)
mask = self.extract(blur)
self.line_detect_possible_demo(mask, center, 20)
except IndexError:
self.src = cv.imread(self.name)
self.src = cv.resize(self.src, dsize=None, fx=0.5, fy=0.5) # 此处可以修改插值方式interpolation
self.src = cv.normalize(self.src, dst=None, alpha=200, beta=10, norm_type=cv.NORM_MINMAX)
blur = cv.pyrMeanShiftFiltering(self.src, 10, 17)
mask = self.extract(blur)
self.line_detect_possible_demo(mask, center, 20)
if __name__ == '__main__':
apparatus = Apparatus('./BONC/1_0555.jpg')
# 读取图片
apparatus.test()
k = cv.waitKey(0)
if k == 27:
cv.destroyAllWindows()
elif k == ord('s'):
cv.imwrite('new.jpg', apparatus.src)
cv.destroyAllWindows()
长指针读数：77.070291
短指针读数：218.896747
由结果可以看出精确度还是挺高的，但是这种方法有三个缺点：
- 当两个指针重合时候不太好处理
- 有时候hough直线检测只能检测出箭头的一条边，这时候就会报错，可以利用图像增强、角点检测和图像梯度来辅助解决，但是效果都不太好
- 计算角度很复杂！！（也可能是我想复杂了，不过这段代码确实花了大量时间）
代码里可能还有很多问题，希望大家多多指出