OpenCV-Python 模板匹配的银行卡号识别

实现流程

1. 对数字模板进行处理，提取出单一数字的图片；

数字模板图片：

单一数字图片示例：

2. 对银行卡图片进行预处理，提取出每组数字图像；

银行卡图片：

一组数字图片示例：

3. 对每组数字图片，提取出单个数字图片，与模板图片进行模板匹配，寻找最优匹配结果，找到对应的数字。

单个数字图片示例：

模板数字处理

数字模板图片：

对模板图片进行二值化处理

#对模板二值化处理
ret , thre_temple = cv2.threshold(temple_gray, 200, 255, cv2.THRESH_BINARY_INV)

二值化后图片：

对二值化后的图片进行轮廓检测

#轮廓检测
temple_contours , hierarchy = cv2.findContours(thre_temple.copy(), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
draw_con_temple = cv2.drawContours(temple, temple_contours, -1, (0, 0, 250), 1)
#print(np.array(temple_contours).shape)
cv_show('draw_con_temple', draw_con_temple)

轮廓检测效果：

确定每个模板数字的图像

根据各个轮廓寻找最小外接矩形，根据最小外接矩形的$y$坐标的大小，对每个数字图片提取并存放在字典的对应序号中。

def get_contours_sort(contours):
    #找到外接矩形，根据y坐标大小
    boundingboxs = [cv2.boundingRect(cnt) for cnt in contours]
    bounding_boxs = sorted(boundingboxs)
    return bounding_boxs

#确定各个数字模板位置
bounding_boxs = get_contours_sort(temple_contours)

#对每个模板提取并存储对应的数字
temple_difits = {}
for (i, c) in enumerate(bounding_boxs):
    (x, y, w, h) = c
    roi = thre_temple[y:y+h , x:x+w ]
    #resize成合适大小
    roi = cv2.resize(roi, (50, 77))
    temple_difits[i] = roi

cv_show('temple_number', temple_difits[0])

单个数字图片示意：

银行卡图片处理

预处理

对银行卡图片进行各种预处理操作，滤除背景影响，同时使得每组数字连在一起；

#礼帽操作，突出更明亮的区域
tophat_goal = cv2.morphologyEx(goal, cv2.MORPH_TOPHAT, rectKernel)
cv_show('credit_card', tophat_goal)

#sobel算子,边缘检测
gradX = cv2.Sobel(tophat_goal, ddepth = cv2.CV_32F, dx = 1, dy = 0)
gradX = np.abs(gradX)
(maxval , minval) = (np.max(gradX) , np.min(gradX))
gradX = (255*(gradX-minval)/(maxval-minval))
gradX = gradX.astype("uint8")
cv_show('sobel', gradX)

#闭操作
close_goal = cv2.morphologyEx(gradX, cv2.MORPH_CLOSE, rectKernel)
cv_show('close', close_goal)

#二值化
ret , thresh_goal = cv2.threshold(close_goal, 0, 255,  cv2.THRESH_OTSU)
cv_show('otsu', thresh_goal)

#闭操作
close2_goal = cv2.morphologyEx(thresh_goal, cv2.MORPH_CLOSE, rectKernel)
cv_show('close2', close_goal)

#二值化
ret , thre_goal = cv2.threshold(close2_goal, 200, 255, cv2.THRESH_BINARY)
cv_show('threshold', thre_goal)

要达到的效果示意：

如上图所示，各组数字连在了一起，同时背景也被滤掉。

确定每组数字

对预处理后的图片进行轮廓检测，根据轮廓外接矩形的长宽比确定每组数字位置。

#轮廓检测
goal_contours , hierarchy = cv2.findContours(thre_goal.copy(), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
draw_con_goal = cv2.drawContours(color_goal.copy(), goal_contours, -1, (0, 0, 250), 1)
#print(np.array(temple_contours).shape)
cv_show('draw_con_goal', draw_con_goal)

goal_boxs = find_goal_part(goal_contours)
num_of_num_part = goal_boxs.shape[0]
#print(goal_boxs)
#存储每个数字片段
goal_difits = {}
color_goal_difits = {}
for i in range(num_of_num_part):
    (x, y, w, h) = goal_boxs[i,:]
    roi = goal[y-5:y+h+5 , x-5:x+w+5 ]
    color_roi = color_goal[y-5:y+h+5 , x-5:x+w+5 ]
    color_goal_difits[i] = color_roi
    goal_difits[i] = roi

def find_goal_part(contours):
    #找到外接矩形，根据长宽比进行选取
    boundingboxs = [cv2.boundingRect(cnt) for cnt in contours]
    rate = []
    for (i , c) in enumerate(boundingboxs):
        (x, y, w, h) = c
        rate.append(w/h)
    rate = np.array(rate)
    number_gath_id = np.argwhere((rate >= 3.55) & (rate <= 3.75))
    boundingboxs = np.array(boundingboxs)
    bounding_boxs =  boundingboxs[number_gath_id]
    bounding_boxs = bounding_boxs[:,0]
    sort_id = np.argsort(bounding_boxs[:,0])
    bounding_boxs = bounding_boxs[sort_id]
    return bounding_boxs

提取出的一组数字图片示例：

轮廓检测每个数字

对提取出的每一组图片进行二值化处理，并进行轮廓检测；

#对数字区域二值化处理
ret , num_part_temple = cv2.threshold(number_part, 200, 255, cv2.THRESH_BINARY)
cv_show('num_part_temple', num_part_temple)

#轮廓检测
num_part_contours , hierarchy = cv2.findContours(num_part_temple.copy(), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
draw_con_num_part = cv2.drawContours(color_num_part.copy(), num_part_contours, -1, (0, 0, 250), 1)
#print(np.array(draw_con_num_part).shape)
#cv_show('draw_con_num_part', draw_con_num_part)

二值化效果：

提取数字组中的数字

根据轮廓的外接矩形的$y$坐标大小按顺序提取每个数字

#确定各个数字位置
num_part_boxs = get_contours_sort(num_part_contours)
num_in_num_part = np.array(num_part_boxs).shape[0]

#对每个数字提取
num_part_difits = {}
for (i, c) in enumerate(num_part_boxs):
    (x, y, w, h) = c
    roi = num_part_temple[y:y+h , x:x+w ]
    #resize成合适大小
    roi = cv2.resize(roi, (50, 77))
    cv_show('roi', roi)

提取出的数字：

模板匹配

对提取的数字进行模板匹配，最佳匹配图片序号即为当前图片数字。

for j in range(10):
    match_result[j] = cv2.matchTemplate(roi, temple_difits[j], cv2.TM_SQDIFF_NORMED)
min_val ,max_val ,min_id ,max_id = cv2.minMaxLoc(match_result)
result.append(min_id[1])

输出结果：

附录

完整代码

#识别银行卡上数字

import cv2
import numpy as np
import matplotlib.pyplot as plt

def cv_show(name,img):
    cv2.imshow(name, img)
    cv2.waitKey(0)
    cv2.destroyAllWindows()
    
def get_contours_sort(contours):
    #找到外接矩形，根据x位置排序
    boundingboxs = [cv2.boundingRect(cnt) for cnt in contours]
    bounding_boxs = sorted(boundingboxs)
    return bounding_boxs

def find_goal_part(contours):
    #找到外接矩形，根据长宽比进行选取
    boundingboxs = [cv2.boundingRect(cnt) for cnt in contours]
    rate = []
    for (i , c) in enumerate(boundingboxs):
        (x, y, w, h) = c
        rate.append(w/h)
    rate = np.array(rate)
    number_gath_id = np.argwhere((rate >= 3.55) & (rate <= 3.75))
    boundingboxs = np.array(boundingboxs)
    bounding_boxs =  boundingboxs[number_gath_id]
    bounding_boxs = bounding_boxs[:,0]
    sort_id = np.argsort(bounding_boxs[:,0])
    bounding_boxs = bounding_boxs[sort_id]
    return bounding_boxs
    

temple = cv2.imread('./orc_credit_num/temple.png' , 1)
temple_gray = cv2.imread('./orc_credit_num/temple.png' , 0)
goal = cv2.imread('orc_credit_num/credit_card_picture.png' , 0)
color_goal = cv2.imread('./orc_credit_num/temple.png' , 1)

cv_show('temple', temple_gray)

#对模板二值化处理
ret , thre_temple = cv2.threshold(temple_gray, 200, 255, cv2.THRESH_BINARY_INV)
cv_show('thre_temple', thre_temple)

#轮廓检测
temple_contours , hierarchy = cv2.findContours(thre_temple.copy(), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
draw_con_temple = cv2.drawContours(temple, temple_contours, -1, (0, 0, 250), 1)
#print(np.array(temple_contours).shape)
cv_show('draw_con_temple', draw_con_temple)

#确定各个数字模板位置
bounding_boxs = get_contours_sort(temple_contours)

#对每个模板提取并存储对应的数字
temple_difits = {}
for (i, c) in enumerate(bounding_boxs):
    (x, y, w, h) = c
    roi = thre_temple[y:y+h , x:x+w ]
    #resize成合适大小
    roi = cv2.resize(roi, (50, 77))
    temple_difits[i] = roi

cv_show('temple_number', temple_difits[0])

#对目标图像进行操作
#初始化卷积核
rectKernel = cv2.getStructuringElement(cv2.MORPH_RECT, (30,15))
sqKernel = cv2.getStructuringElement(cv2.MORPH_RECT, (5,5))

cv_show('credit_card', goal)

#礼帽操作，突出更明亮的区域
tophat_goal = cv2.morphologyEx(goal, cv2.MORPH_TOPHAT, rectKernel)
cv_show('credit_card', tophat_goal)

#sobel算子,边缘检测
gradX = cv2.Sobel(tophat_goal, ddepth = cv2.CV_32F, dx = 1, dy = 0)
gradX = np.abs(gradX)
(maxval , minval) = (np.max(gradX) , np.min(gradX))
gradX = (255*(gradX-minval)/(maxval-minval))
gradX = gradX.astype("uint8")
cv_show('sobel', gradX)

#闭操作
close_goal = cv2.morphologyEx(gradX, cv2.MORPH_CLOSE, rectKernel)
cv_show('close', close_goal)

#二值化
ret , thresh_goal = cv2.threshold(close_goal, 0, 255,  cv2.THRESH_OTSU)
cv_show('otsu', thresh_goal)

#闭操作
close2_goal = cv2.morphologyEx(thresh_goal, cv2.MORPH_CLOSE, rectKernel)
cv_show('close2', close_goal)

#二值化
ret , thre_goal = cv2.threshold(close2_goal, 200, 255, cv2.THRESH_BINARY)
cv_show('threshold', thre_goal)

#轮廓检测
goal_contours , hierarchy = cv2.findContours(thre_goal.copy(), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
draw_con_goal = cv2.drawContours(color_goal.copy(), goal_contours, -1, (0, 0, 250), 1)
#print(np.array(temple_contours).shape)
cv_show('draw_con_goal', draw_con_goal)

goal_boxs = find_goal_part(goal_contours)
num_of_num_part = goal_boxs.shape[0]
#print(goal_boxs)
#存储每个数字片段
goal_difits = {}
color_goal_difits = {}
for i in range(num_of_num_part):
    (x, y, w, h) = goal_boxs[i,:]
    roi = goal[y-5:y+h+5 , x-5:x+w+5 ]
    color_roi = color_goal[y-5:y+h+5 , x-5:x+w+5 ]
    color_goal_difits[i] = color_roi
    goal_difits[i] = roi

cv_show('goal_difits', goal_difits[3])
result = []
for k in range(num_of_num_part):
    #开始识别数字
    number_part = goal_difits[k]
    color_num_part = color_goal_difits[k]

    #对数字区域二值化处理
    ret , num_part_temple = cv2.threshold(number_part, 200, 255, cv2.THRESH_BINARY)
    cv_show('num_part_temple', num_part_temple)

    #轮廓检测
    num_part_contours , hierarchy = cv2.findContours(num_part_temple.copy(), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
    draw_con_num_part = cv2.drawContours(color_num_part.copy(), num_part_contours, -1, (0, 0, 250), 1)
    #print(np.array(draw_con_num_part).shape)
    #cv_show('draw_con_num_part', draw_con_num_part)

    #确定各个数字位置
    num_part_boxs = get_contours_sort(num_part_contours)
    num_in_num_part = np.array(num_part_boxs).shape[0]

    #对每个数字提取
    num_part_difits = {}
    for (i, c) in enumerate(num_part_boxs):
        (x, y, w, h) = c
        roi = num_part_temple[y:y+h , x:x+w ]
        #resize成合适大小
        roi = cv2.resize(roi, (50, 77))
        cv_show('roi', roi)
        match_result = np.zeros((10,1))
        for j in range(10):
            match_result[j] = cv2.matchTemplate(roi, temple_difits[j], cv2.TM_SQDIFF_NORMED)
        min_val ,max_val ,min_id ,max_id = cv2.minMaxLoc(match_result)
        result.append(min_id[1])
    
print(result)