see

Why we included this project

This is the reference implementation for the AAAI 2018 paper on semi-supervised end-to-end scene text recognition, so it is a good place to start if you want to reproduce a published method or build on it instead of writing everything from scratch. The code shows how a CNN-based detection and recognition pipeline can be trained with semi-supervised learning, which helps when labeled scene-text data is hard to come by. One honest caveat: it is built on Chainer and targets CUDA 8/9-era tooling, so expect to spend some time updating the environment before it runs on a modern stack. Teams doing OCR on photos or street signs can still learn a lot from the architecture and training approach, even if they end up porting the ideas to another framework.