赞
踩
最终用这个keras版本训练CRNN:https://codeload.github.com/xiaofengShi/CHINESE-OCR/zip/master
详细过程看我的OneNote
app端:https://github.com/chineseocr/chineseocr
训练ocr端:https://codeload.github.com/chineseocr/chineseocr/zip/master
最近基于mxnet的cnocr模型挺小的,据说坑太多,有空尝试一下:https://github.com/breezedeus/cnocr
基于pytorch的超级小的开源中文OCR人气很高,支持竖排文字,不过没有训练代码,有空尝试一下:https://github.com/ouyanghuiyu/chineseocr_lite
ctcloss编译 未成功
根据:https://www.freesion.com/article/5487477118/
模型转换:
pytorch ocr 转keras ocrpython .\pytorch_to_keras.py -weights_path models/ocr-english.pth -output_path models/ocr-english.h5
遇到的bug:
keras——ValueError: You are trying to load a weight file containing 12 layers into a model with 2 layers.
model=load_model('./model/my_model_weights.h5')
改为:
model.load_weights('test.h5',by_name=True)
Copyright © 2003-2013 www.wpsshop.cn 版权所有,并保留所有权利。