从表格图像中提取单个字段以使用 OCR 进行 Excel

2022-01-09 00:00:00 python opencv computer-vision image-processing ocr

问题描述

我已经扫描了具有如下图所示表格的图像:

I have scanned images which have tables as shown in this image:

我正在尝试分别提取每个框并执行 OCR，但是当我尝试检测水平线和垂直线然后检测框时，它会返回以下图像:

I am trying to extract each box separately and perform OCR but when I try to detect horizontal and vertical lines and then detect boxes it's returning the following image:

当我尝试执行其他转换来检测文本(腐蚀和扩张)时，仍然会出现一些剩余的线条，如下所示:

And when I try to perform other transformations to detect text (erode and dilate) some remains of lines are still coming along with text like below:

我无法检测到仅用于执行 OCR 的文本，并且未生成正确的边界框，如下所示:

I cannot detect text only to perform OCR and proper bounding boxes aren't being generated like below:

我无法使用实线得到清晰分隔的框，我已经在一个用paint(如下所示)编辑的图像上尝试了这个以添加数字并且它有效.

I cannot get clearly separated boxes using real lines, I've tried this on an image that was edited in paint(as shown below) to add digits and it works.

我不知道我做错了哪一部分，但如果有什么我应该尝试或更改/添加我的问题，请告诉我.

I don't know which part I'm doing wrong but if there's anything I should try or maybe change/add in my question please please tell me.

#Loading all required libraries 
%pylab inline
import cv2
import numpy as np 
import pandas as pd
import pytesseract
import matplotlib.pyplot as plt
import statistics
from time import sleep
import random

img = cv2.imread('images/scan1.jpg',0)

# for adding border to an image
img1= cv2.copyMakeBorder(img,50,50,50,50,cv2.BORDER_CONSTANT,value=[255,255])

# Thresholding the image
(thresh, th3) = cv2.threshold(img1, 255, 255,cv2.THRESH_BINARY|cv2.THRESH_OTSU)

# to flip image pixel values
th3 = 255-th3

# initialize kernels for table boundaries detections
if(th3.shape[0]<1000):
    ver = np.array([[1],
               [1],
               [1],
               [1],
               [1],
               [1],
               [1]])
    hor = np.array([[1,1,1,1,1,1]])

else:
    ver = np.array([[1],
               [1],
               [1],
               [1],
               [1],
               [1],
               [1],
               [1],
               [1],
               [1],
               [1],
               [1],
               [1],
               [1],
               [1],
               [1],
               [1],
               [1],
               [1]])
    hor = np.array([[1,1,1,1,1,1,1,1,1,1,1,1,1,1,1]])




# to detect vertical lines of table borders
img_temp1 = cv2.erode(th3, ver, iterations=3)
verticle_lines_img = cv2.dilate(img_temp1, ver, iterations=3)

# to detect horizontal lines of table borders
img_hor = cv2.erode(th3, hor, iterations=3)
hor_lines_img = cv2.dilate(img_hor, hor, iterations=4)

# adding horizontal and vertical lines
hor_ver = cv2.add(hor_lines_img,verticle_lines_img)

hor_ver = 255-hor_ver

# subtracting table borders from image
temp = cv2.subtract(th3,hor_ver)

temp = 255-temp

#Doing xor operation for erasing table boundaries
tt = cv2.bitwise_xor(img1,temp)

iii = cv2.bitwise_not(tt)

tt1=iii.copy()

#kernel initialization
ver1 = np.array([[1,1],
               [1,1],
               [1,1],
               [1,1],
               [1,1],
               [1,1],
               [1,1],
               [1,1],
               [1,1]])
hor1 = np.array([[1,1,1,1,1,1,1,1,1,1],
               [1,1,1,1,1,1,1,1,1,1]])

#morphological operation
temp1 = cv2.erode(tt1, ver1, iterations=2)
verticle_lines_img1 = cv2.dilate(temp1, ver1, iterations=1)

temp12 = cv2.erode(tt1, hor1, iterations=1)
hor_lines_img2 = cv2.dilate(temp12, hor1, iterations=1)

# doing or operation for detecting only text part and removing rest all
hor_ver = cv2.add(hor_lines_img2,verticle_lines_img1)
dim1 = (hor_ver.shape[1],hor_ver.shape[0])
dim = (hor_ver.shape[1]*2,hor_ver.shape[0]*2)

# resizing image to its double size to increase the text size
resized = cv2.resize(hor_ver, dim, interpolation = cv2.INTER_AREA)

#bitwise not operation for fliping the pixel values so as to apply morphological operation such as dilation and erode
want = cv2.bitwise_not(resized)

if(want.shape[0]<1000):
    kernel1 = np.array([[1,1,1]])
    kernel2 = np.array([[1,1],
                        [1,1]])
    kernel3 = np.array([[1,0,1],[0,1,0],
                       [1,0,1]])
else:
    kernel1 = np.array([[1,1,1,1,1,1]])
    kernel2 = np.array([[1,1,1,1,1],
                        [1,1,1,1,1],
                        [1,1,1,1,1],
                        [1,1,1,1,1]])

tt1 = cv2.dilate(want,kernel1,iterations=2)

# getting image back to its original size
resized1 = cv2.resize(tt1, dim1, interpolation = cv2.INTER_AREA)

# Find contours for image, which will detect all the boxes
contours1, hierarchy1 = cv2.findContours(resized1, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)

#function to sort contours by its x-axis (top to bottom)
def sort_contours(cnts, method="left-to-right"):
    # initialize the reverse flag and sort index
    reverse = False
    i = 0

    # handle if we need to sort in reverse
    if method == "right-to-left" or method == "bottom-to-top":
        reverse = True

    # handle if we are sorting against the y-coordinate rather than
    # the x-coordinate of the bounding box
    if method == "top-to-bottom" or method == "bottom-to-top":
        i = 1

    # construct the list of bounding boxes and sort them from top to
    # bottom
    boundingBoxes = [cv2.boundingRect(c) for c in cnts]
    (cnts, boundingBoxes) = zip(*sorted(zip(cnts, boundingBoxes),
        key=lambda b:b[1][i], reverse=reverse))

    # return the list of sorted contours and bounding boxes
    return (cnts, boundingBoxes)


#sorting contours by calling fuction
(cnts, boundingBoxes) = sort_contours(contours1, method="top-to-bottom")

#storing value of all bouding box height
heightlist=[]
for i in range(len(boundingBoxes)):
    heightlist.append(boundingBoxes[i][3])

#sorting height values
heightlist.sort()

sportion = int(.5*len(heightlist))
eportion = int(0.05*len(heightlist))

#taking 50% to 95% values of heights and calculate their mean 
#this will neglect small bounding box which are basically noise 
try:
    medianheight = statistics.mean(heightlist[-sportion:-eportion])
except:
    medianheight = statistics.mean(heightlist[-sportion:-2])

#keeping bounding box which are having height more then 70% of the mean height and deleting all those value where 
# ratio of width to height is less then 0.9
box =[]
imag = iii.copy()
for i in range(len(cnts)):    
    cnt = cnts[i]
    x,y,w,h = cv2.boundingRect(cnt)
    if(h>=.7*medianheight and w/h > 0.9):
        image = cv2.rectangle(imag,(x+4,y-2),(x+w-5,y+h),(0,255,0),1)
        box.append([x,y,w,h])
    # to show image

###Now we have badly detected boxes image as shown

解决方案

你在正确的轨道上.这是您的方法的延续，稍作修改.这个想法是:

You're on the right track. Here's a continuation of your approach with slight modifications. The idea is:

  1. 获取二值图像.加载图像，转灰度，大津阈值.

  1. Obtain binary image. Load image, convert to grayscale, and Otsu's threshold.

删除所有字符文本轮廓.我们创建一个矩形内核并执行打开以仅保留水平/垂直线.这将有效地使文本变成微小的噪声，因此我们找到轮廓并使用轮廓区域进行过滤以去除它们.

Remove all character text contours. We create a rectangular kernel and perform opening to only keep the horizontal/vertical lines. This will effectively make the text into tiny noise so we find contours and filter using contour area to remove them.

修复水平/垂直线并提取每个 ROI. 我们变形接近修复和断线并平滑表格.从这里我们使用带有 top-to-bottom 参数的 imutils.sort_contours() 对框域轮廓进行排序.接下来我们找到轮廓并使用轮廓区域进行过滤，然后提取每个 ROI.

Repair horizontal/vertical lines and extract each ROI. We morph close to fix and broken lines and smooth the table. From here we sort the box field contours using imutils.sort_contours() with the top-to-bottom parameter. Next we find contours and filter using contour area then extract each ROI.

<小时>

这是每个框字段和提取的 ROI 的可视化

Here's a visualization of each box field and the extracted ROI

代码

import cv2
import numpy as np
from imutils import contours

# Load image, grayscale, Otsu's threshold
image = cv2.imread('1.jpg')
original = image.copy()
gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
thresh = cv2.threshold(gray, 0, 255, cv2.THRESH_BINARY_INV + cv2.THRESH_OTSU)[1]

# Remove text characters with morph open and contour filtering
kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (3,3))
opening = cv2.morphologyEx(thresh, cv2.MORPH_OPEN, kernel, iterations=1)
cnts = cv2.findContours(opening, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
cnts = cnts[0] if len(cnts) == 2 else cnts[1]
for c in cnts:
    area = cv2.contourArea(c)
    if area < 500:
        cv2.drawContours(opening, [c], -1, (0,0,0), -1)

# Repair table lines, sort contours, and extract ROI
close = 255 - cv2.morphologyEx(opening, cv2.MORPH_CLOSE, kernel, iterations=1)
cnts = cv2.findContours(close, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
cnts = cnts[0] if len(cnts) == 2 else cnts[1]
(cnts, _) = contours.sort_contours(cnts, method="top-to-bottom")
for c in cnts:
    area = cv2.contourArea(c)
    if area < 25000:
        x,y,w,h = cv2.boundingRect(c)
        cv2.rectangle(image, (x, y), (x + w, y + h), (36,255,12), -1)
        ROI = original[y:y+h, x:x+w]

        # Visualization
        cv2.imshow('image', image)
        cv2.imshow('ROI', ROI)
        cv2.waitKey(20)

cv2.imshow('opening', opening)
cv2.imshow('close', close)
cv2.imshow('image', image)
cv2.waitKey()

相关文章