At the preprocessing stage, the text is rotated to the horizontal position in the scanned image. 

After that, it is cleaned of background, noise and artifacts. 

It doesn’t matter if the image was scanned in color (24-bit RGB) or grayscale mode. It is converted to black and white (1-bit) during binarization. 

After this procedure, the resulting monochrome image is decomposed into a binary code consisting only of zeros and ones (where 0 is white and 1 is black). 

The next step is to zone-out the text into columns, rows, paragraphs and tables. 

The separately recognized letter “A” or “O” may be part of the English word or Russian.  Therefore, at this stage the document’s language is often defined.  

During character recognition, the simplest algorithms analyze each pixel contained in black and white images. They compare each character with a database of known fonts. The result of the recognition is the character that is the closest match.  

More advanced algorithms break down each character into elements, such as lines, strokes, bends and angles between them, and analyze the joints of these elements. 

The ability to connect dictionaries at the post-processing stage increases the probability of correct character recognition and eliminates the possibility that non-existent words may appear in the text. You can set certain rules in these algorithms.  For example, you have the scanned envelope. There is a field on it containing the postal code. It means there can only be digits from 0 to 9. Knowing this, the algorithm will eliminate the appearance of the letter “O” instead of the number “0”, the letter “B” instead of the number “8” or the Latin letter “I” instead of the number “1”. Telephone numbers, car plates and VAT numbers are also easily formalized. 

By now, thanks to machine learning, OCR algorithms can recognize complex fonts as well as handwritten text.