Remove background text and noise from an image using image processing with OpenCV

Question

I have these images

For which I want to remove the text in the background. Only the captcha characters should remain(i.e K6PwKA, YabVzu). The task is to identify these characters later using tesseract.

This is what I have tried, but it isn't giving much good accuracy.

import cv2
import pytesseract

pytesseract.pytesseract.tesseract_cmd = r"C:\Users\HPO2KOR\AppData\Local\Tesseract-OCR\tesseract.exe"
img = cv2.imread("untitled.png")
gray_image = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
gray_filtered = cv2.inRange(gray_image, 0, 75)
cv2.imwrite("cleaned.png", gray_filtered)

How can I improve the same?

Note : I tried all the suggestion that I was getting for this question and none of them worked for me.

EDIT : According to Elias, I tried finding the color of the captcha text using photoshop by converting it to grayscale which came out to be somewhere in between [100, 105]. I then threshold the image based on this range. But the result which I got did not give satisfactory result from tesseract.

gray_filtered = cv2.inRange(gray_image, 100, 105)
cv2.imwrite("cleaned.png", gray_filtered)
gray_inv = ~gray_filtered
cv2.imwrite("cleaned.png", gray_inv)
data = pytesseract.image_to_string(gray_inv, lang='eng')

Output :

'KEP wKA'

Result :

EDIT 2 :

def get_text(img_name):
    lower = (100, 100, 100)
    upper = (104, 104, 104) 
    img = cv2.imread(img_name)
    img_rgb_inrange = cv2.inRange(img, lower, upper)
    neg_rgb_image = ~img_rgb_inrange
    cv2.imwrite('neg_img_rgb_inrange.png', neg_rgb_image)
    data = pytesseract.image_to_string(neg_rgb_image, lang='eng')
    return data

gives :

and the text as

GXuMuUZ

Is there any way to soften it a little

Answer 1

Didn't try , but this might work. step 1: use ps to find out what color the captcha characters are. For excample, "YabVzu" is (128,128,128),

Step 2: Use pillow's method getdata()/getcolor(), it will return a sequence which contain the colour of every pixel.

then ,we project every item in the sequence to the original captcha image.

hence we know the positon of every pixel in the image.

Step 3: find all pixels whose colour with the most approximate values to (128,128,128). You may set a threshold to control the accuracy. this step return another sequence. Lets annotate it as Seq a

Step 4: generate a picture with the very same height and width as the original one. plot every pixel in [Seq a] in the very excat position in the picture. Here,we will get a cleaned training items

Step 5: Use a Keras project to break the code. And the precission should be over 72%.