If the task is to recognize handwritten letters , then which method / algorithm of machine learning is the most relevant and gives more accurate results in this area at the current moment? For example, is it better to use neural networks, or is it possible to completely do without machine learning methods of machine learning?

Now I have the task of recognizing handwritten letters (Cyrillic). Letters will be recognized separately (not words entirely). At the entrance I will submit tables with various handwritten letters of the Russian alphabet. Need help in choosing the most optimal algorithm / method for recognizing handwritten characters.

  • one
    In my opinion, the majority now use convolutional neural networks (Convolutional Neural Networks) for such tasks. - MaxU

2 answers 2

When it comes to image recognition, the answer immediately appears - the convolutional neural network. Your case is no exception. I must say at once that there are many ready-made solutions with the already prepared architecture of the National Assembly, so you don’t need to write anything yourself. If C ++ or Python, I advise Tensorflow . To understand how Convolutional NAs work, read my answer to a similar question.

    No neural networks, especially convolutional ones, are needed.
    Take Tesseract , train in handwriting and use.
    Tesseract itself uses NA, but not convolutional :-)

    Most recently, I created a project — Tesseract is used to identify charboxes, then Tesseract tries to recognize what it has revealed, in parallel, the same charboxes are trying to recognize an algorithm based on OpenCV kNN .

    For what is not recognized, there is a GUI in which you can select the charbox and specify manually what character it is. As a result, data is generated for tesseract training (what he could not recognize himself and was entered manually), and data for training kNN (what tesseract recognized and what was manually entered). After> 10 workouts per character, it starts to be fairly consistently recognized by both systems.