CAPTCHA from http://login.weibo.cn/login/

finshied.

Programing Language: Python
Library: Tensorflow-1.0 + scikit-learn-0.18 + Pillow + opencv-python

use some computer vision algorithm to clean the peper noise and noise line, vertical projection to split the word, and CNN to train the dataset.

now is unavailable, see the backup at captchas_backup.zip.

cd ./spliter
mkdir dataset
python __main__.py

You can check the results in ./trainer/training_set.zip

cd ./trainer/
unzip training_set.zip

cd ./trainer/
python format_dataset.py
python train.py

cd ./recognizer
python recognizer test1.png test2.png test3.png test4.png

The accuracy is about 60%.