By using Python (numpy/sklearn/opencv) I'm trying to process images with hand-written digits, and after all my main problem turns to be connected digits. For example:

If I crop the digits by hand I get expected output, by taking advantage from NIST dataset, and sklearn SVC classifier, but segmenting the digits seems like impossible task to me. Does anyone know of programmatic approach (like library or similar) for this task, regardless programming language, but preferable in Python?