使用 opencv 从图像中移除 flash 眩光

Question

I'm working on extracting the the electricity meter number from a photos of a watt-hour meter using OpenCV python. I have the photos cropped/detected using Yolov4 like this: The original Image

Then I do some operations on it including, making it gray, blur, thresholding, and then closing. Photos without glare are working fine. But the image like the one above is not.

`#gray scaling gray_img = cv.cvtColor(img, cv.COLOR_RGB2GRAY)

blur_img = cv.GaussianBlur(gray_img, (5,5), 0)

rect_kern = cv.getStructuringElement(cv.MORPH_RECT, (5,5))

ret, thresh = cv.threshold(blur_img, 0, 255, cv.THRESH_OTSU | cv.THRESH_BINARY_INV)

closing = cv.morphologyEx(thresh, cv.MORPH_CLOSE, rect_kern, iterations=2)

dilation = cv.dilate(thresh, rect_kern, iterations = 1)

#finding the contours

contours, hierarchy = cv.findContours(closing, cv.RETR_TREE, cv.CHAIN_APPROX_SIMPLE)

sorted_contours = sorted(contours, key=lambda ctr: cv.boundingRect(ctr)[0])`

Then after that, I compute the ratio of the width and height of every contour to find the numbers and pass them to the ocr, but that's not my purpose of asking this question. The result of the above image and the above code is the following.

The final image before reading with the ocr

As you can see around the number 9 and 1 there is the glare which resulted in not being able to read the numbers. Is there anyone to solve this issue. I'm open to any kind of advice, even probably using something other than OpenCV. Btw I'm using tesseract as an ocr.

Answer 1

Glare is in fact destroying any information locally, in an unrecoverable way.

A solution could be to use an OCR engine that allows "don't know" pixels, but AFAIK Tesseract does not support that.

The damage is lessened if you turn the bright pixels to dark before other processing.