How to extract numbers from image with reflections and artifacts?

Hello.

I have a series of photos of the seven-digit display (below please find the example of such photo). I want to apply OCR to extract the information from each consecutive frame. Generally, the methods works quite fine provided that the image is distinct. Hovewer, at the preprocessing stage there is a need to binarize the image. The problem lies in a fact that there are some reflections in the image. I spent a significant amount of time and tried a lot of combinations and various functions (e.g. adaptive thresholding, histograms) to obtain the best possible performance. Is there any reasonable method for obtaining a nice set of digits without artifacts? Unfortunately there is no way to repeat the experiments in better conditions and remove the reflections at acqusition stage.

Thank you kindly in advance for any useful suggestions.

[DGM]

Are all the images taken from a fixed camera position? If so, then there may be a way to either develop a template or at least map the reflections in order to help compensate.

Answer 1

[Image Analyst]

@DGM is right. It's better to fix your camera image acquisition parameters, if you can, than to try to fix up a bad image. If all the images are taken from the same position it would be better because you could then simply use a template to determine if the mean intensity in a region is above or below a threshold. That would be a little more work than simply using the OCR function though. Also get control over the light and possibly use polarizers to knock down reflections.

If you can't improve the images, like they're taken from users in the wild with smartphones, they try adapthisteq to try to make all the digits fairly contrasty. Then try OCR. If it's still not good, then try preprocessing the image to take out the background or make it more uniform.

For what it's worth, I'm attaching a 7 segment display demo I did years ago.