[Python] How to repair obscured blocks in a QR code image

I have many RGB images that contain a printed sheet of paper with a QR code in an outdoor setting. Because of the bright sun interfering with image capture, about 20% of images are unreadable:



I'm using magick in R to handle the image manipulation, and then the Python package pyzbar (which wraps zbar) to do the detection.

I can use image_threshold to find all the pixels that are within the 95% quantile and force them to pure black, which fixes about half my images:



But some of them remain unreadable, like this one. I can see with my human eyes that I need to fill in some of the anchor points in the upper left, so I mocked this up in MS Paint:



With that manual manipulation, this image is now easily read. Is there any way to do this kind of repair automatically? I don't mind translating code from Python or the ImageMagick CLI, so R-only answers aren't necessary.

My general approach:

library(magick)
library(reticulate)
pyzbar <- import("pyzbar.pyzbar")

magick_to_numpy <- function(img) {
round(255 * as.numeric(magick::image_data(img, "rgb")))
}

image_read("testfile.jpg") %>%
image_threshold("black", "95%") %>%
magick_to_numpy() %>%
pyzbar$decode()


Usual result:

list()​

Desired result:

[[1]]
Decoded(data=b'W TRR C6 T2', type='QRCODE',
rect=Rect(left=176, top=221, width=373, height=333),
polygon=[Point(x=176, y=226), Point(x=202, y=554),
Point(x=549, y=544), Point(x=524, y=221)])​

Continue reading...