Photographing Chinese-Address OCR implemented using CTPN+CTC+Address Correction.
This is a project of the 2018 Deecamp 25th group (DRPRG). Thanks to my team members!
这个是 2018年 Deecamp 25组 (深度受限抠图小组) 的项目，非常非常的感谢每一位队友！
Our Demo: https://www.bilibili.com/video/av30081208
Our Wechat Program (微信小程序)： OCRdeecamp
Reference Paper: Detecting Text in Natural Image with Connectionist Text Proposal Network https://arxiv.org/pdf/1609.03605.pdf
Reference Code: https://github.com/YCG09/chinese_ocr (Thanks Yang Chenguang)
Text Detection : CTPN (https://arxiv.org/pdf/1609.03605.pdf)
Text Recognition: CTC+DenseNet
Address Judgment: Light GBM or textgrocery (https://github.com/2shou/TextGrocery)
Address Correction: Fuzzy matching based on address library
You can simpely use demo_final.py for inference. Input a picture and output the Chinese address string.
Communicate between server and wechat program using flask ocr_whole.py
Text detection using CTPN, then text recognition using Densenet
Address correction using fuzyy-matching based on address library
If you want to know more details about CTPN codes, please check https://github.com/eragonruan/text-detection-ctpn
Some files of wechat program (微信小程序的一些文件)
In our dataset, the accuracy of exactly correct is 83% ,accuracy of edit distance less than 3 is 97%. This result is better than Baidu and Tencent's general OCR.
Our program has a high accuracy at identifying very fuzzy multi-line addresses.
If you want to know more details, please read 'ChineseAddress_OCR_Report.pdf'(Chinese). If you want to read details in English, please contact me.