tricks - wanghaisheng/awesome-ocr GitHub Wiki

No. I tried ocropy with hidden nodes of 200 and found, as far as I estimate, that it began to learn one char by forgetting another. I am training clstm against 3877 classes of Chinese/Japanese characters with hidden node = 800. After 150000 iteration, it keeps 3.8-5% error rate. See clstm section.

https://github.com/tleyden/open-ocr Run your own OCR-as-a-Service using Tesseract and Docker http://www.openocr.net/