使用传统的计算机视觉方法定位图像中的车牌,参考了部分网上的文章,实际定位效果对于我目前使用的网上的图片来说还可以。实测发现对于车身本身是蓝色、或是车牌本身上方有明显边缘的情况这类图片定位效果较差。纯练手项目,仅供参考。代码中imagePreProcess对某些图片定位率相比于imagePreProcess2做预处理的效果要好。后续可以尝试做一个如果imagePreProcess2识别无效后使用imagePreProcess再处理,或者加上阈值自适应打分的机制优化。目前对于我做的练手项目来说足够了。
注意:以下代码是参考了网上的一些文章后,按照自己的思路写的,定位效果尚可。参考的文章有:python-opencv实战:车牌识别(一):精度还不错的车牌定位_基于阈值分割的车牌定位识别-CSDN博客
https://www.cnblogs.com/fyunaru/p/12083856.html
import cv2 as cv
import numpy as np
import matplotlib.pyplot as plt
#过滤矩形的参数
minRectW = 100
minRectH = 50
#判断车牌颜色的参数
#一般情况下,蓝色车牌H分量的值通常都在115附近徘徊
# S分量和V分量因光照不同而差异较大(opencv中H分量的取值范围是0到179,而不是图像学中的0到360;S分量和V分量的取值范围是到255)
deltaH = 15
hsvLower = np.array([115 - deltaH,60,60])
hsvUpper = np.array([115 + deltaH,255,255])
#灰度拉伸
def grayScaleStretch(img):
maxGray = float(img.max())
minGray = float(img.min())
for i in range(img.shape[0]):
for j in range(img.shape[1]):
img[i,j] = 255 / (maxGray - minGray) * (img[i,j] - minGray)
return img
#图像二值化
def image2Binary(img):
#选取灰度最大最小值的中间值
maxGray = float(img.max())
minGray = float(img.min())
threshold = (minGray + maxGray) / 2
ret,bin = cv.threshold(img, threshold, 255, cv.THRESH_BINARY)
return bin
#图像预处理
def imagePreProcess(img):
#转换为灰度图
imgGray = cv.cvtColor(img, cv.COLOR_BGR2GRAY)
#灰度拉伸
imgGray = grayScaleStretch(imgGray)
#plt.imshow(imgGray, cmap='gray')
kernel = cv.getStructuringElement(cv.MORPH_ELLIPSE, (3,3))
#做开运算
imgOpen = cv.morphologyEx(imgGray, cv.MORPH_OPEN, kernel)
#plt.imshow(imgOpen, cmap='gray')
#获得差分图
imgDiff = cv.absdiff(imgGray, imgOpen)
#plt.imshow(imgDiff, cmap='gray')
imgDiff = cv.GaussianBlur(imgDiff, (3,3), 5)
#plt.imshow(imgDiff, cmap='gray')
#图像二值化
imgBinary = image2Binary(imgDiff)
#plt.imshow(imgBinary, cmap='gray')
cannyEdges = cv.Canny(imgBinary, 127, 200)
#plt.imshow(cannyEdges, cmap='gray')
#对Canny检测边缘结果做处理
kernel = np.ones((3,3), np.uint8)
imgOut = cv.morphologyEx(cannyEdges, cv.MORPH_CLOSE, kernel)
imgOut = cv.dilate(imgOut, kernel, iterations=1)
imgOut = cv.morphologyEx(imgOut, cv.MORPH_OPEN, kernel)
#imgOut = cv.erode(imgOut, kernel, iterations=1)
imgOut = cv.morphologyEx(imgOut, cv.MORPH_CLOSE, kernel)
#plt.imshow(imgOut, cmap='gray')
return imgOut
#图像预处理2 - 对于某些
def imagePreProcess2(img):
imgGray = cv.cvtColor(img, cv.COLOR_BGR2GRAY)
#灰度拉伸
imgGray = grayScaleStretch(imgGray)
imgGray = cv.GaussianBlur(imgGray, (3,3), 5)
#进行边缘检测
cannyEdges = cv.Canny(imgGray, 180, 230)
#二值化
imgBinary = image2Binary(cannyEdges)
#plt.imshow(imgBinary, cmap='gray')
#先做闭运算再做开运算
kernel = np.ones((3,3), np.uint8)
imgOut = cv.morphologyEx(imgBinary, cv.MORPH_CLOSE, kernel)
imgOut = cv.morphologyEx(imgOut, cv.MORPH_OPEN, kernel)
imgOut = cv.absdiff(imgBinary, imgOut)
imgOut = cv.morphologyEx(imgOut, cv.MORPH_CLOSE, kernel)
imgOut = cv.dilate(imgOut, kernel, iterations=1)
plt.imshow(imgOut, cmap='gray')
return imgOut
#debug
def printHSV(hsvSrc):
for i in range(hsvSrc.shape[0]):
for j in range(hsvSrc.shape[1]):
(h,s,v) = hsvSrc[i][j]
print(h,s,v)
#定位车牌
def locate_plate(imgProcessing, imgOriginal):
contours,hierarchy = cv.findContours(imgProcessing, cv.RETR_EXTERNAL, cv.CHAIN_APPROX_SIMPLE)
carPlateCandidates = []
for contour in contours:
(x,y,w,h) = cv.boundingRect(contour)
#过滤掉一些小的矩形
if (w < minRectW or h < minRectH):
continue
#cv.rectangle(imgOriginal, (int(x), int(y)), (int(x + w),int(y + h)), (0,255,0), 2)
carPlateCandidates.append([int(x),int(y),int(x + w),int(y + h)])
#plt.imshow(imgOriginal[:,:,::-1])
maxMean = 0
target = []
target_mask = []
#依次检查候选车牌列表,用HSV颜色空间判别是否是车牌
for candidate in carPlateCandidates:
(x0,y0,x1,y1) = candidate
candidateROI = imgOriginal[y0:y1,x0:x1]
hsvROI = cv.cvtColor(candidateROI, cv.COLOR_BGR2HSV)
mask = cv.inRange(hsvROI, hsvLower, hsvUpper)
#print(mask)
#plt.imshow(mask, cmap='gray')
#使用均值找出蓝色最多的区域
mean = cv.mean(mask)
#print(mean)
if mean[0] > maxMean:
maxMean = mean[0]
target = candidate
target_mask = mask
#对target的范围进行缩小,找出蓝色刚开始和结束的坐标
print(target_mask)
nonZeroPoints = cv.findNonZero(target_mask)
#print(nonZeroPoints)
sortByX = np.sort(nonZeroPoints, axis=0)
xMin = sortByX[0][0][0]
xMax = sortByX[-1][0][0]
print(sortByX)
sortByY = np.sort(nonZeroPoints, axis=1)
yMin = sortByY[0][0][1]
yMax = sortByY[-1][0][1]
print(sortByY)
print("X min:" + str(xMin) + " X max:" + str(xMax) + " Y min:" + str(yMin) + " Y max:" + str(yMax))
(x0,y0,x1,y1) = target
print("Original:" + str(x0) + "," + str(y0) + "," + str(x1) + "," + str(y1))
#target = (x0 + xMin, y0 + yMin, x0 + (xMax - xMin), y0 + yMax - yMin)
target = [x0 + xMin, y0 + yMin, x0 + xMax, y0 + yMax]
return target
#读取图像
imgCarPlate = cv.imread("../../SampleImages/carplate/carplate_chongqing.jpg", cv.IMREAD_COLOR)
#plt.imshow(imgCarPlate[:,:,::-1])
img4locate = imagePreProcess2(imgCarPlate)
target = locate_plate(img4locate, imgCarPlate)
(x0,y0,x1,y1) = target
cv.rectangle(imgCarPlate, (x0,y0), (x1,y1), (0,255,0), 2)
plt.imshow(imgCarPlate[:,:,::-1])
成功的例子:
不太成功的例子(轮廓检测的不太好,并且轮廓中蓝色的值过早出现,可以优化判断为连续的蓝色而不是零散的蓝色)
失败的例子(没能检测出小轮廓,车身本身为蓝色,替换为imagePreProcess后能够成功):