I have to recognize digits within an image from video stream and there are several more things, that should make recognition easier:
1) it is fixed font 6×8, all symbols are equal width
2) I know exact positions of digits, they are always rectangular, are not rotated/sqewed/scaled, but there may be some distortions because of air transmission glitch.
3) It is only digits and .
4) digit background is semi black (50% opaque)
I’ve tried tesseract v2 and v3, but .NET wrappers aren’t perfect and recognition error was very large, even if I trained with custom font, as far as I understand that is because of small resolution.
I’ve made very simple algorithm by my self by turning image to black and white and counting matching pixels between original font image and image from stream, it performs better than tesseract, but I hink more sophisticated algorithm would do better.
I’ve tried to train AForge using ActivationNetwork with BackPropagationLearning and it fails to converge(this article first part, as long as I don’t need scaling and several fonts http://www.codeproject.com/Articles/11285/Neural-Network-OCR, as I understand code in article is for older version of AForge), bad thing is, that this project is not supported anymore, forum is closed and google groups as I understand too.
I know there’s OpenCV port to .NET, as far as I see, it has different network approaches than AForge, so questiton is which approach would fit best.
So is there any .NET framework to help me at this, and if it supports more than one neural network implementations, which implementation would fit best?
For fixed size fonts at a fixed magnification, you can probably get away with a less-sophisticated OCR approach based on template matching. See here for an example of how to do template matching using OpenCV (not .NET, but hopefully enough to get you started.) The basic idea is that you create a template for each digit, then try matching all templates at your target location, choosing the one with the highest match score. Because you know where the digits are located, you can search over a very small area for each digit. For more information on the theory behind template-matching, see this wiki article on Cross-correlation.
This is actually the basis for simplified OCR applications (usually for recognizing special OCR fonts, like the SEMI standard fonts used for printing serial numbers on silicon wafers.) The production-grade algorithms can also support tolerance for scaling, rotation and translation, but the underlying techniques are pretty much the same.