效果展示
流程介绍
- 先把图片进行处理，提取数字部分
- 将四个数字分割，并进行模板匹配
- 处理数字和区域，标注在原图上
步骤
进行图片处理
- 图片太大，需要进行缩放，等会标注的时候需要放大回来。
- 进行一次顶帽操作， 扩大特征
- 用sobel进行边缘检测，这里用的是x方向的检测，然后进行归一化处理，把值控制在255内
sobelX = cv.Sobel(top_hat, cv.CV_32F, 1, 0)
- cv_32F目标是32位有符号浮点数，因为sobel算子是用的右边减去左边，有可能为负数，这样保留正负数，再进行绝对值操作
sobelX = np.absolute(sobelX)
(min, max) = (np.min(top_hat), np.max(top_hat))
sobelX = (255 * ((sobelX - min)/ (max - min)))
sobelX = sobelX.astype('uint8')
- 进行一步闭合操作，去除噪点
- 进行二值化操作
- 进行闭操作，连接成块
- 进行边缘检测
- 过滤选择区域
- 进行数字区域分割提取
- 进行数字内容识别
def find_num(ori_path):
card_path = ori_path
card_img = cv.imread(card_path)
card_img = cv.resize(card_img, None, fx=0.7, fy=0.7)
# 灰度化
card_gray = cv.cvtColor(card_img, cv.COLOR_BGR2GRAY)
kernel = cv.getStructuringElement(cv.MORPH_RECT, (5, 5))
rectKernel = cv.getStructuringElement(cv.MORPH_RECT, (13, 3))
top_hat = cv.morphologyEx(card_gray, cv.MORPH_TOPHAT, kernel)
# 算子sobel
sobelX = cv.Sobel(top_hat, cv.CV_32F, 1, 0)
sobelX = np.absolute(sobelX)
(min, max) = (np.min(top_hat), np.max(top_hat))
sobelX = (255 * ((sobelX - min)/ (max - min)))
sobelX = sobelX.astype('uint8')
# 二值化
card_close = cv.morphologyEx(sobelX, cv.MORPH_CLOSE, rectKernel)
ret, threshold = cv.threshold(card_close, 0, 255, cv.THRESH_BINARY | cv.THRESH_OTSU)
card_gray = cv.morphologyEx(threshold, cv.MORPH_CLOSE, kernel, iterations=1)
contours, hierachy = cv.findContours(card_gray, cv.RETR_TREE, cv.CHAIN_APPROX_SIMPLE)
# cv.drawContours(card_img, contours, -1, (0, 255, 0), 4)
list = []
for cont in contours:
# 画矩形
x, y, w, h = cv.boundingRect(cont)
rate = w/h
if 2.7 < rate < 3.5:
if (50 < w < 95) and (10 < h < 40):
# rect = cv.rectangle(card_img, (x, y), (x+w, y+h), (0, 255, 0), 1)
# cv.imshow('card1', card_img[y: y + h, x: x + w])
list.append(([x, y, w, h], card_img[y: y + h, x: x + w]))
# 从左到右排序
list = sorted(list, key= lambda a: a[0][0])
num_list = []
for cont, img in list:
four_list = ""
img1 = cv.cvtColor(img, cv.COLOR_BGR2GRAY)
ret, img1 = cv.threshold(img1, 0, 255, cv.THRESH_BINARY | cv.THRESH_OTSU)
contours1, hierachy = cv.findContours(img1, cv.RETR_EXTERNAL, cv.CHAIN_APPROX_SIMPLE)
# cv.drawContours(img, contours1, -1, (0, 255, 0), 1)
temp_list = []
for cont1 in contours1:
# 绘制矩形
x, y, w, h = cv.boundingRect(cont1)
rect = cv.rectangle(img1, (x, y), (x+w, y+h), (0, 255, 0), 1)
rect1 = img1[y: y + h , x : x + w ]
temp_list.append(([x, y, w, h], rect1))
temp_list = sorted(temp_list, key= lambda a: a[0][0])
for i, img2 in temp_list:
four_list += (match_img(img2))
num_list.append(four_list)
print(num_list)
return list, num_list
模板的处理
- 直接用常规操作进行分割
- 写入本地数据
def split_template():
reference_path = './data/reference.png'
img_base = cv.imread(reference_path)
img_gray = cv.cvtColor(img_base, cv.COLOR_BGR2GRAY)
ret, threshold = cv.threshold(img_gray, 0, 255, cv.THRESH_BINARY_INV)
contours, hierachy = cv.findContours(threshold, cv.RETR_EXTERNAL, cv.CHAIN_APPROX_SIMPLE)
for i, cont in enumerate(contours):
x, y, w, h = cv.boundingRect(cont)
cv.imwrite('template/%s.png'%i, threshold[y- 5: y + h + 5, x - 5: x + w + 5])
匹配模板
- 因为上面导出的图片名称为
- 所以数字应该用 9- 下标，应该注意下。
import os
def match_img(img):
max = 0
max_index = 0
for i,name in enumerate(os.listdir("./template")):
img_bg = cv.imread('./template/%s'%name, 0)
img = cv.resize(img, (img_bg.shape[1], img_bg.shape[0]))
result = cv.matchTemplate(img_bg, img, cv.TM_CCOEFF_NORMED)
if result > max:
max = result
max_index = 9 - i
return str(max_index)
展示匹配效果
- 这里需要传入原始图片， 图片列表， 数字列表
- 遍历标注即可
def draw_card(img, list, num_list):
# 传入坐标list， 数字list (list, numList)
for i, cont in enumerate(list):
x, y, w, h = cont[0]
x = int(x/0.7); y = int(y/0.7); w = int(w/0.7); h = int(h/0.7)
print(x, y, w, h)
cv.rectangle(img, (x, y), (x + w, y + h), (0, 255, 0), 2)
num_str = num_list[i]
cv.putText(img, num_str, (x, y - 10), cv.FONT_HERSHEY_COMPLEX, 1, (0, 255, 0), 2, cv.LINE_AA)
cv.imshow('card', img)
cv.waitKey(0)
cv.destroyAllWindows()
调用主函数
def main():
ori_path = './data/card3.png'
split_template()
list, num_list = find_num(ori_path)
draw_card(cv.imread(ori_path), list, num_list)
main()
效果
缺陷
- 这里有张银色的会有反光效果，处理起来不是很好，而且下面的英文也和数字的比例很像。
- 有找到处理方案的同学，可以留言评论
资源获取
完整代码
import numpy as np
import cv2 as cv
from PIL import Image
import os
def match_img(img):
max = 0
max_index = 0
for i,name in enumerate(os.listdir("./template")):
img_bg = cv.imread('./template/%s'%name, 0)
img = cv.resize(img, (img_bg.shape[1], img_bg.shape[0]))
result = cv.matchTemplate(img_bg, img, cv.TM_CCOEFF_NORMED)
if result > max:
max = result
max_index = 9 - i
return str(max_index)
def draw_card(img, list, num_list):
# 传入坐标list， 数字list (list, numList)
for i, cont in enumerate(list):
x, y, w, h = cont[0]
x = int(x/0.7); y = int(y/0.7); w = int(w/0.7); h = int(h/0.7)
print(x, y, w, h)
cv.rectangle(img, (x, y), (x + w, y + h), (0, 255, 0), 2)
num_str = num_list[i]
cv.putText(img, num_str, (x, y - 10), cv.FONT_HERSHEY_COMPLEX, 1, (0, 255, 0), 2, cv.LINE_AA)
cv.imshow('card', img)
cv.waitKey(0)
cv.destroyAllWindows()
def find_num(ori_path):
card_path = ori_path
card_img = cv.imread(card_path)
card_img = cv.resize(card_img, None, fx=0.7, fy=0.7)
# 灰度化
card_gray = cv.cvtColor(card_img, cv.COLOR_BGR2GRAY)
kernel = cv.getStructuringElement(cv.MORPH_RECT, (5, 5))
rectKernel = cv.getStructuringElement(cv.MORPH_RECT, (13, 3))
top_hat = cv.morphologyEx(card_gray, cv.MORPH_TOPHAT, kernel)
# 算子sobel
sobelX = cv.Sobel(top_hat, cv.CV_32F, 1, 0)
sobelX = np.absolute(sobelX)
(min, max) = (np.min(top_hat), np.max(top_hat))
sobelX = (255 * ((sobelX - min)/ (max - min)))
sobelX = sobelX.astype('uint8')
# 二值化
card_close = cv.morphologyEx(sobelX, cv.MORPH_CLOSE, rectKernel)
ret, threshold = cv.threshold(card_close, 0, 255, cv.THRESH_BINARY | cv.THRESH_OTSU)
card_gray = cv.morphologyEx(threshold, cv.MORPH_CLOSE, kernel, iterations=1)
contours, hierachy = cv.findContours(card_gray, cv.RETR_TREE, cv.CHAIN_APPROX_SIMPLE)
# cv.drawContours(card_img, contours, -1, (0, 255, 0), 4)
list = []
for cont in contours:
# 画矩形
x, y, w, h = cv.boundingRect(cont)
rate = w/h
if 2.7 < rate < 3.5:
if (50 < w < 95) and (10 < h < 40):
# rect = cv.rectangle(card_img, (x, y), (x+w, y+h), (0, 255, 0), 1)
# cv.imshow('card1', card_img[y: y + h, x: x + w])
list.append(([x, y, w, h], card_img[y: y + h, x: x + w]))
# 从左到右排序
list = sorted(list, key= lambda a: a[0][0])
num_list = []
for cont, img in list:
four_list = ""
img1 = cv.cvtColor(img, cv.COLOR_BGR2GRAY)
ret, img1 = cv.threshold(img1, 0, 255, cv.THRESH_BINARY | cv.THRESH_OTSU)
contours1, hierachy = cv.findContours(img1, cv.RETR_EXTERNAL, cv.CHAIN_APPROX_SIMPLE)
# cv.drawContours(img, contours1, -1, (0, 255, 0), 1)
temp_list = []
for cont1 in contours1:
# 绘制矩形
x, y, w, h = cv.boundingRect(cont1)
rect = cv.rectangle(img1, (x, y), (x+w, y+h), (0, 255, 0), 1)
rect1 = img1[y: y + h , x : x + w ]
temp_list.append(([x, y, w, h], rect1))
temp_list = sorted(temp_list, key= lambda a: a[0][0])
for i, img2 in temp_list:
four_list += (match_img(img2))
num_list.append(four_list)
print(num_list)
return list, num_list
def split_template():
reference_path = './data/reference.png'
img_base = cv.imread(reference_path)
img_gray = cv.cvtColor(img_base, cv.COLOR_BGR2GRAY)
ret, threshold = cv.threshold(img_gray, 0, 255, cv.THRESH_BINARY_INV)
contours, hierachy = cv.findContours(threshold, cv.RETR_EXTERNAL, cv.CHAIN_APPROX_SIMPLE)
for i, cont in enumerate(contours):
x, y, w, h = cv.boundingRect(cont)
cv.imwrite('template/%s.png'%i, threshold[y- 5: y + h + 5, x - 5: x + w + 5])
def main():
ori_path = './data/card4.png'
split_template()
list, num_list = find_num(ori_path)
draw_card(cv.imread(ori_path), list, num_list)
main()
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
