python

Weibo.cn

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

The Captcha Image

Status

finshied.

Enviorment

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

Technique

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.

Steps

1.Download some captchas.

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

2.Split the letters from every captcha.

cd ./spliter
mkdir dataset
python __main__.py

3. Recognize the letters by human.

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

cd ./trainer/
unzip training_set.zip

4. Train the dataset.

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

5. Test the recognizer

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

About the Accuracy

The accuracy is about 60%.