CODEWITHSUNDEEP
c#androidopencvimage-processingxamarin.android
Kacper
Kacper

Reputation: 622

OpenCV grayscale ignore red color

I am making app for recognize digits (OCR) so I need to prepare imamge for it. There is no problem when i take photo to blue, green, yellow or other color but red digits become so gray after grayscale in OpenCV and these digits became unable to recognize.

Original image: enter image description here

Image after grayscale (yellow and red digits): enter image description here

Image after threshold:

enter image description here

As you can see, after it red digits are gone.

Here is fragment of code, that I use:

mat.ConvertTo(mat, CvType.Cv8uc1);
Imgproc.CvtColor(mat, mat, Imgproc.ColorBgr2gray);
Imgproc.Threshold(mat, mat, 127, 255, Imgproc.ThreshBinary);

Any solutions?

Upvotes: 2

Views: 3517

Answers (2)

bfris
bfris

Reputation: 5805

@Jeru Luke's solution should be fairly robust for a wide range of input images. But if you need raw speed, you might think about a simple brightness/contrast operation followed by global thresholding.

If you use brightness and contrast, which is computationally cheap, you can make the background become all black and then use global thresholding to get a nice binarized image.

Photo editors (Photoshop, Gimp, etc.) often use brightness/contrast scale of ±127. The mathematical formala for simultaneously adding brightness (b) and contrast (c) is

img = (1 + c/127)*img + (b-c)

If you have access to mat from C#, then you can use the cv.mat.convertTo function:

cv.Mat.convertTo( OutputArray, cv.CV_8U, 1+c/127, b-c)

For your image I used b = -45 and c = +45

Then convert to grayscale and binarize (I used threshold of 50 on your image)

Update

The OP was tagged for C#. But many of us use Python. In Python, we have no access to Mat. However, we can use the cv2.addWeighted function which does:

dst = src1*alpha + src2*beta + gamma

If we set beta = 0, then this becomes equivalent to cv.Mat.convertTo scaling. This seems to be faster than doing matrix operations in Numpy. Numpy is a little slower because we have to do some extra stuff to handle overflow.

Upvotes: 2

Jeru Luke
Jeru Luke

Reputation: 21203

As I mentioned in the comments you can perform Otsu threshold to each of the color channels R, G, B.

Otsu threshold of Blue channel:

enter image description here

Otsu threshold of Green channel:

enter image description here

Otsu threshold of Red channel:

enter image description here

Finally I added all the above to get the following result:

enter image description here

I only used the following functions:

  1. cv2.threshold()

  2. cv2.add()

UPDATE

Code

import os
import cv2
import numpy as np

#--- performs Otsu threshold ---
def threshold(img, st):
    ret, thresh = cv2.threshold(img, 0, 255, cv2.THRESH_BINARY | cv2.THRESH_OTSU)
    cv2.imwrite(os.path.join(path, 'res_' + str(st) + '.jpg'), thresh) 
    return  thresh

path = r'C:\Users\Desktop'
filename = 'digits.jpg'

img = cv2.imread(os.path.join(path, filename))
img = cv2.resize(img, (0, 0), fx = 0.5, fy = 0.5)   #--- resized the image because it was to big 
cv2.imshow('Original', img)

#--- see each of the channels individually ---
cv2.imshow('b', img[:,:,0])
cv2.imshow('g', img[:,:,1])
cv2.imshow('r', img[:,:,2])

m1 = threshold(img[:,:,0], 1)   #--- threshold on blue channel
m2 = threshold(img[:,:,1], 2)   #--- threshold on green channel
m3 = threshold(img[:,:,2], 3)   #--- threshold on red channel

#--- adding up all the results above ---
res = cv2.add(m1, cv2.add(m2, m3))

cv2.imshow('res', res)
cv2.imwrite(os.path.join(path, 'res.jpg'), res)

cv2.waitKey()
cv2.destroyAllWindows()

Upvotes: 4

Related Questions