1 背景

利用opencv识别数码管，采用传统的穿线法，前提是利用深度学习目标检测的方法，在机柜中将数码管区域切割出来，然后再对切割出来的区域进行识别，判断数字是多少。切割出来的形状如图所示：

数码管各段表示为

2 识别思路

先对数码管进行灰度化和二值化，将数字变为255，背景变为0，然后利用穿线法，对abcdefg7个区域依次穿线，判断是否有255的值，有则表示该区域高亮，最后结合7个区域的高亮信息，综合判断数字是多少

a b c d e f g result

0 √ √ √ √ √ √ 63

1 √ √ 6

2 √ √ √ √ √ 91

3 √ √ √ √ √ 79

4 √ √ √ √ 102

5 √ √ √ √ √ 109

6 √ √ √ √ √ √ 125

7 √ √ √ 7

8 √ √ √ √ √ √ √ 127

9 √ √ √ √ √ 103

3 灰度化

opencv有自带的灰度化函数cv2.cvtColor()，但是在使用过程中发现，对于一些整体亮度低的数码管灰度化后，会丢失数字信息，看不出来哪段亮，考虑到数码管都是红色的，红色通道的数据最重要，因此设计了一个自己灰度化的函数 def tomygray(image):

height = image.shape[0]

width = image.shape[1]

gray = np.zeros((height, width, 1), np.uint8)

for i in range(height):

for j in range(width):

# pixel = max(image[i,j][0], image[i,j][1], image[i,j][2])

pixel = 0.0 * image[i, j][0] + 0.0 * image[i, j][1] + 1 * image[i, j][2]

gray[i, j] = pixel

return gray

4 二值化

opencv有多种二值化的方法，主要包括固定阈值和自适应阈值的方法，具体介绍可看《OpenCV之阈值化操作总结》

自适应阈值主要适用于一张图片中亮度不一样的情况，而对于我们的数码管来说，由于大小很小，基本上没有亮度变化，因此使用固定阈值的方法即可。函数原型如下 ret, dst = cv2.threshold(src, thresh, maxval, type)

函数最重要的部分是thresh值的设置，由于不同图片的数码管亮度不同，不可以选择同一个阈值，需要分别计算每张图片的固定阈值，计算阈值有很多方法，我用到的方法有以下两种

4.1 统计直方图

统计直方图中像素的分布情况，根据数量最多的像素值来设置一个阈值(下边的参数都是调试效果比较好的值，自己可根据具体情况来设置)

hist = cv2.calcHist([image_gray], [0], None, [256], [0,256])

#plt.hist(hist.ravel(), 256, [0,256])

#plt.savefig(filename + '_hist.png')

#plt.show()

min_val, max_val, min_index, max_index = cv2.minMaxLoc(hist)

ret, image_bin = cv2.threshold(image_gray, int(max_index[1])-7, 255,

cv2.THRESH_BINARY)

4.2 计算平均值

计算灰度图的平均像素值，根据平均值设定阈值 mean,stddev = cv2.meanStdDev(image_gray)

ret, image_bin = cv2.threshold(image_gray, meanvalue + 65, 255,

cv2.THRESH_BINARY)

5 穿线法

得到二值化的图像后，将图像进行分割，切成一个一个的数字，然后每个都用穿线法来判断值是多少

def TubeIdentification(filename, num, image):

tube = 0

tubo_roi = [

[image.shape[0] * 0/3, image.shape[0] * 1/3, image.shape[1] * 1/2,

image.shape[1] * 1/2],

[image.shape[0] * 1/3, image.shape[0] * 1/3, image.shape[1] * 2/3,

image.shape[1] - 1 ],

[image.shape[0] * 2/3, image.shape[0] * 2/3, image.shape[1] * 2/3,

image.shape[1] - 1 ],

[image.shape[0] * 2/3, image.shape[0] -1 , image.shape[1] * 1/2,

image.shape[1] * 1/2],

[image.shape[0] * 2/3, image.shape[0] * 2/3, image.shape[1] * 0/3,

image.shape[1] * 1/3],

[image.shape[0] * 1/3, image.shape[0] * 1/3, image.shape[1] * 0/3,

image.shape[1] * 1/3],

[image.shape[0] * 1/3, image.shape[0] * 2/3, image.shape[1] * 1/2,

image.shape[1] * 1/2]]

i = 0

while(i < 7):

if(Iswhite(image, int(tubo_roi[i][0]), int(tubo_roi[i][1]),

int(tubo_roi[i][2]),int(tubo_roi[i][3]))):

tube = tube + pow(2,i)

cv2.line(image, ( int(tubo_roi[i][3]),int(tubo_roi[i][1])),

(int(tubo_roi[i][2]), int(tubo_roi[i][0])),

(255,0,0), 1)

i += 1

if(tube==63):

onenumber = 0

elif(tube==6):

onenumber = 1

elif(tube==91):

onenumber = 2

elif(tube==79):

onenumber = 3

elif(tube==102 or tube==110):

#110是因为有干扰情况

onenumber = 4

elif(tube==109):

onenumber = 5

elif(tube==125):

onenumber = 6

elif(tube==7):

onenumber = 7

elif(tube==127):

onenumber = 8

elif(tube==103):

onenumber = 9

else:

onenumber = -1

cv2.imwrite(filename + '_' + str(num) + '_' + str(onenumber) + '.png', image)

return onenumber

def Iswhite(image, row_start, row_end, col_start, col_end):

white_num = 0

j=row_start

i=col_start

while(j <= row_end):

while(i <= col_end):

if(image[j][i] == 255):

white_num+=1

i+=1

j+=1

i=col_start

#print('white num is',white_num)

if(white_num >= 5):

return True

else:

return False

6 识别主程序 def digitalrec(image):

filename = str(image).split('.jpg', 1)[0]

image_org = cv2.imread(image)

height = image_org.shape[0]

width = image_org.shape[1]

#transe image to gray

#image_gray = cv2.cvtColor(image_org, cv2.COLOR_RGB2GRAY)

image_gray = tomygray(image_org)

cv2.imwrite(filename + '_gray.png',image_gray)

meanvalue = image_gray.mean()

if meanvalue >= 200:

hist = cv2.calcHist([image_gray], [0], None, [256], [0,256])

#plt.hist(hist.ravel(), 256, [0,256])

#plt.savefig(filename + '_hist.png')

#plt.show()

min_val, max_val, min_index, max_index = cv2.minMaxLoc(hist)

ret, image_bin = cv2.threshold(image_gray, int(max_index[1])-7, 255,

cv2.THRESH_BINARY)

else:

mean,stddev = cv2.meanStdDev(image_gray)

ret, image_bin = cv2.threshold(image_gray, meanvalue + 65, 255,

cv2.THRESH_BINARY)

#image_bin = cv2.adaptiveThreshold(image_gray, 255,

# cv2.ADAPTIVE_THRESH_GAUSSIAN_C,

# cv2.THRESH_BINARY, 11,

# 0)

x, y, w, h = cv2.boundingRect(image_bin)

image_bin = image_bin[max(y-5,0) : h+10, max(x-5,0) : w+10]

cv2.imwrite(filename + '_bin.png',image_bin)

#split number and identify it

num = 0

result = ''

while True:

if(num < 3):

roi = image_bin[0: height, int(width / 3 * num):

int(width / 3 * (num + 1))]

onenumber = TubeIdentification(filename, num, roi)

if(onenumber == -1):

result += '0'

else:

result += str(onenumber)

num += 1

else:

break

print('picture of %s detect result is %s'%(filename,result))

return result

7 识别效果

8 展望

上边介绍的方法可以实现数码管的识别，但是由于有很多对像素的操作，比较耗时，平均识别一张图片需要2s左右，而且使用传统的方法来识别数码管，涉及到很多参数的设置与调试，鲁棒性不强，尤其是在灰度化和二值化的时候，参数设置很关键。

最好的解决方法是可以利用深度学习的方法来识别，提高识别率，有时间的话会尝试新的方法来解决。