Python + OpenCV: OCR Image Segmentation

Question

I am trying to do OCR from this toy example of Receipts. Using Python 2.7 and OpenCV 3.1.

Grayscale + Blur + External Edge Detection + Segmentation of each area in the Receipts (for example "Category" to see later which one is marked -in this case cash-).

I find complicated when the image is "skewed" to be able to properly transform and then "automatically" segment each segment of the receipts.

Example:

Any suggestion?

The code below is an example to get until the edge detection, but when the receipt is like the first image. My issue is not the Image to text. Is the pre-processing of the image.

Any help more than appreciated! :)

import os;
os.chdir() # Put your own directory

import cv2 
import numpy as np

image = cv2.imread("Rent-Receipt.jpg", cv2.IMREAD_GRAYSCALE)

blurred = cv2.GaussianBlur(image, (5, 5), 0)

#blurred  = cv2.bilateralFilter(gray,9,75,75)

# apply Canny Edge Detection
edged = cv2.Canny(blurred, 0, 20)

#Find external contour

(_,contours, _) = cv2.findContours(edged, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_NONE)

Answer 1

A great tutorial on the first step you described is available at pyimagesearch (and they have great tutorials in general)

In short, as described by Ella, you would have to use cv2.CHAIN_APPROX_SIMPLE . A slightly more robust method would be to use cv2.RETR_LIST instead of cv2.RETR_EXTERNAL and then sort the areas, as it should decently work even in white backgrounds/if the page inscribes a bigger shape in the background, etc.

Coming to the second part of your question, a good way to segment the characters would be to use the Maximally stable extremal region extractor available in OpenCV. A complete implementation in CPP is available here in a project I was helping out in recently. The Python implementation would go along the lines of (Code below works for OpenCV 3.0+. For the OpenCV 2.x syntax, check it up online)

import cv2

img = cv2.imread('test.jpg')
mser = cv2.MSER_create()

#Resize the image so that MSER can work better
img = cv2.resize(img, (img.shape[1]*2, img.shape[0]*2))

gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
vis = img.copy()

regions = mser.detectRegions(gray)
hulls = [cv2.convexHull(p.reshape(-1, 1, 2)) for p in regions[0]]
cv2.polylines(vis, hulls, 1, (0,255,0)) 

cv2.namedWindow('img', 0)
cv2.imshow('img', vis)
while(cv2.waitKey()!=ord('q')):
    continue
cv2.destroyAllWindows()

This gives the output as

Now, to eliminate the false positives, you can simply cycle through the points in hulls, and calculate the perimeter (sum of distance between all adjacent points in hulls[i], where hulls[i] is a list of all points in one convexHull). If the perimeter is too large, classify it as not a character.

The diagnol lines across the image are coming because the border of the image is black. that can simply be removed by adding the following line as soon as the image is read (below line 7)

img = img[5:-5,5:-5,:]

which gives the output

Answer 2

The option on the top of my head requires the extractions of 4 corners of the skewed image. This is done by using cv2.CHAIN_APPROX_SIMPLE instead of cv2.CHAIN_APPROX_NONE when finding contours. Afterwards, you could use cv2.approxPolyDP and hopefully remain with the 4 corners of the receipt (If all your images are like this one then there is no reason why it shouldn't work).

Now use cv2.findHomography and cv2.wardPerspective to rectify the image according to source points which are the 4 points extracted from the skewed image and destination points that should form a rectangle, for example the full image dimensions.

Here you could find code samples and more information: OpenCV-Geometric Transformations of Images

Also this answer may be useful - SO - Detect and fix text skew

EDIT: Corrected the second chain approx to cv2.CHAIN_APPROX_NONE .

Answer 3

Preprocessing the image by converting the desired text in the foreground to black while turning unwanted background to white can help to improve OCR accuracy. In addition, removing the horizontal and vertical lines can improve results. Here's the preprocessed image after removing unwanted noise such as the horizontal/vertical lines. Note the removed border and table lines

import cv2

# Load in image, convert to grayscale, and threshold
image = cv2.imread('1.jpg')
gray = cv2.cvtColor(image,cv2.COLOR_BGR2GRAY)
thresh = cv2.threshold(gray, 0, 255, cv2.THRESH_BINARY_INV + cv2.THRESH_OTSU)[1]

# Find and remove horizontal lines
horizontal_kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (35,2))
detect_horizontal = cv2.morphologyEx(thresh, cv2.MORPH_OPEN, horizontal_kernel, iterations=2)
cnts = cv2.findContours(detect_horizontal, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
cnts = cnts[0] if len(cnts) == 2 else cnts[1]
for c in cnts:
    cv2.drawContours(thresh, [c], -1, (0,0,0), 3)

# Find and remove vertical lines
vertical_kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (1,35))
detect_vertical = cv2.morphologyEx(thresh, cv2.MORPH_OPEN, vertical_kernel, iterations=2)
cnts = cv2.findContours(detect_vertical, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
cnts = cnts[0] if len(cnts) == 2 else cnts[1]
for c in cnts:
    cv2.drawContours(thresh, [c], -1, (0,0,0), 3)

# Mask out unwanted areas for result
result = cv2.bitwise_and(image,image,mask=thresh)
result[thresh==0] = (255,255,255)

cv2.imshow('thresh', thresh)
cv2.imshow('result', result)
cv2.waitKey()

Answer 4

Try using Stroke Width Transform. Python 3 implementation of the algorithm is present here at SWTloc

Install the Library

pip install swtloc

Run the swttransform

from swtloc import SWTLocalizer
from swtloc.utils import imgshowN, imgshow

swtl = SWTLocalizer()
imgpath = ...
swtl.swttransform(imgpaths=imgpath, text_mode = 'lb_df', gs_blurr=False ,
                  minrsw = 3, maxrsw = 10, max_angledev = np.pi/3)
mgshowN([swtl.orig_img, swtl.swt_mat, swtl.swt_labelled3C],
        ['Original Image', 'Stroke Width Transform', 'Connected Components'])

Run the Grouping of texts

respacket = swtl.get_grouped(lookup_radii_multiplier=.8, sw_ratio=2,
                 cl_deviat=[13,13,13], ht_ratio=2, 
                 ar_ratio=4, ang_deviat=30)

grouped_labels = respacket[0]
grouped_bubblebbox = respacket[1]
grouped_annot_bubble = respacket[2]

imgshowN([swtl.orig_img, grouped_annot_bubble],
        ['Original', 'Grouped Bubble BBox Annotation'])

There are multiple parameters in the of the swttransform function and get_grouped function that you can play around with to get the desired results.

Full Disclosure : I am the author of this library