zeusees / HyperOCR

HyperOCR

A Simple light-weight text detection implementation based on CTPN with mobilenet1.0 backbone.

Startup

python demo_ctpn.py --image textx.jpg --prefix model/rpn1 --epoch 80 --gpu 0 --vis

Demo Image

Training Tips

1. Execute script init.sh(init.bat on Windows) to initialize project.
2. Download pretrained model from here into model folder.
3. Download dataset from baidu yun. This dataset is already prepared by @eragonruan to fit CTPN.
4. Unzip the dataset downloaded to 'VOCdevkit' folder, and set both default.root_path and default.dataset_path in rcnn/config.py to '<somewhere>/VOCdevkit/VOC2007'. You can also change other hyperparams in rcnn/config.py.
5. Run python train_ctpn.py to train. Run python train_ctpn.py --gpus '0' --rpn_lr 0.01 --no_flip 0 to train model on gpu 0 + with learning rate 0.01 and with flip data augmentation.

Tips

• This is an experimental , maybe cannot achieve good effect.
• The code based on https://github.com/chinakook/CTPN.mxnet

TODO

• new implementation based on PSENET

• Text region feature extractor for template based ocr.

• OCR model.