Opencv Image transformation for Tesseract OCR

0
votes

I have following Image which I want to feed to tesseract to detect the text:

Input Image:

enter image description here

I am processing this image using OTSU transformation: the code is as follows:

import cv2
import numpy as np
from matplotlib import pyplot as plt
import glob


for img in glob.glob("/home/image.jpg"):

    cv_img=cv2.imread(img,0)    
    #img = cv2.imread('1.jpg',0)
    cv_img = cv2.medianBlur(cv_img,5)

    ret,th1 = cv2.threshold(cv_img,127,255,cv2.THRESH_BINARY)
    th2 = cv2.adaptiveThreshold(cv_img,255,cv2.ADAPTIVE_THRESH_MEAN_C,\
            cv2.THRESH_BINARY,11,2)
    th3 = cv2.adaptiveThreshold(cv_img,255,cv2.ADAPTIVE_THRESH_GAUSSIAN_C,\
            cv2.THRESH_BINARY,11,2)


    cv2.imwrite('OTSU.jpg', th3)

The output Image that I am getting after this transform is:

enter image description here

Here are my conditions: The major hurdle that I am facing is the white lines which appear randomly on the image. The actual image may have more numbers than 1993 but the number of white lines interfering image may increase, I want to clean this image files so that they are ready for Tesseract OCR.

I have tried Canny edge detection on the orignal Image which seems to give the outlines

Canny edge detector Output: enter image description here

How do I clean the input image to get rid of the white lines overlapping on the text? My aim is to run it through Tesseract OCR.

3
python-3.xopencvpython-tesseract

3 Answers

0
votes

I think you should look into to Morphological Transformations in opencv. The documentation can be found here.

Made a small piece of code with a starting point from your script where I have used erosion and dilation: 

import cv2
import numpy as np

cv_img = cv2.imread('1993.jpg', 0)


cv_img = cv2.medianBlur(cv_img, 5)

ret, th1 = cv2.threshold(cv_img,127,255,cv2.THRESH_BINARY)
th2 = cv2.adaptiveThreshold(cv_img,255,cv2.ADAPTIVE_THRESH_MEAN_C, cv2.THRESH_BINARY,11,2)
th3 = cv2.adaptiveThreshold(cv_img,255,cv2.ADAPTIVE_THRESH_GAUSSIAN_C, cv2.THRESH_BINARY,11,2)

kernel_erosion = np.ones((3,3), np.uint8)
kernel_dilation = np.ones((1,1), np.uint8)
erosion = cv2.erode(th2, kernel_erosion, iterations=1)
dilation = cv2.dilate(erosion, kernel_dilation, iterations=1)

cv2.imwrite('morph.jpg', dilation)

You can play around with different kernels or different transformations. This is the output I got:

Erosion and dilation

0
votes

Try using CLAHE before you threshold the image. This is what I tried:

import cv2
import numpy as np

image = cv2.imread("numbers.jpg")

gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)

clahe = cv2.createCLAHE(clipLimit=10, tileGridSize=(5, 5))
gray = clahe.apply(gray)
ret, thresh = cv2.threshold(v, 140, 150, cv2.THRESH_BINARY_INV)
morph = cv2.morphologyEx(thresh, cv2.MORPH_OPEN, np.ones((5,5), np.uint8), 
                        iterations=1)
morph = cv2.morphologyEx(morph, cv2.MORPH_CLOSE, np.ones((3,1), np.uint8), 
                         iterations=1)

Result:

Result

You can replace the threshold type with cv2.THRESH_OTSU and it should work without any issues. Try playing around with the parameters and I'm sure you can make it to work with all your images. Cheers!

0
votes

And one more result with cv2.ximgproc.niBlackThreshold:

enter image description here