基于这位作者，原理请查看链接的文章 http://www.bugcode.cn/break_captcha.html
经过一番郁闷的环境搭建过程，遇到各种奇葩问题（主要是 opencv），浪费了半个小时，我决定基于 docker 构建环境 这里主要是快速部署上的操作
获得 docker 镜像 docker pull yoanlin/opencv-python3-tensorflow
测试一下环境
# python3
Python 3.6.2 (default, Sep 13 2017, 14:26:54)
[GCC 4.9.2] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import cv2
>>> cv2.__version__
'3.2.0'
>>>
成功输出版本号，说明环境正常
cd
wget https://s3-us-west-2.amazonaws.com/mlif-example-code/solving_captchas_code_examples.zip
apt-get install unzip
unzip solving_captchas_code_examples.zip code
cd code
pip3 install -r requirements.txt
pip3 install --upgrade h5py
apt-get update
apt-get install libgtk2.0-dev pkg-config
a.利用 opencv生成单个字符的图片
python3 extract_single_letters_from_captchas.py
文件夹extracted_letter_images下会生成以各个字符命名的图片文件夹
# cd extracted_letter_images
# ls
2 4 6 8 A C E G J L N Q S U W Y
3 5 7 9 B D F H K M P R T V X Z
# cd 2
# ls
000001.png 000202.png 000403.png 000604.png 000805.png 001006.png
000002.png 000203.png 000404.png 000605.png 000806.png 001007.png
000003.png 000204.png 000405.png 000606.png 000807.png 001008.png
000004.png 000205.png 000406.png 000607.png 000808.png 001009.png
000005.png 000206.png 000407.png 000608.png 000809.png 001010.png
cd ../..
b.开始训练（风在吼马在叫，风扇在咆哮） python3 train_model.py 有个小 waring，不影响运行，先忽略
2018-01-19 07:29:15.745077: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use AVX2 instructions, but these are available on your machine and could speed up CPU computations.
2018-01-19 07:29:15.745083: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use FMA instructions, but these are available on your machine and could speed up CPU computations.
29058/29058 [==============================] - 25s 872us/step - loss: 0.1935 - acc: 0.9531 - val_loss: 0.0205 - val_acc: 0.9945
Epoch 2/10
29058/29058 [==============================] - 26s 882us/step - loss: 0.0136 - acc: 0.9963 - val_loss: 0.0176 - val_acc: 0.9940
Epoch 3/10
29058/29058 [==============================] - 28s 974us/step - loss: 0.0069 - acc: 0.9980 - val_loss: 0.0058 - val_acc: 0.9983
Epoch 4/10
29058/29058 [==============================] - 26s 894us/step - loss: 0.0039 - acc: 0.9988 - val_loss: 0.0119 - val_acc: 0.9968
Epoch 5/10
29058/29058 [==============================] - 24s 843us/step - loss: 0.0059 - acc: 0.9983 - val_loss: 0.0085 - val_acc: 0.9977
Epoch 6/10
29058/29058 [==============================] - 25s 847us/step - loss: 0.0034 - acc: 0.9990 - val_loss: 0.0068 - val_acc: 0.9978
Epoch 7/10
29058/29058 [==============================] - 25s 853us/step - loss: 0.0120 - acc: 0.9979 - val_loss: 0.0151 - val_acc: 0.9958
Epoch 8/10
29058/29058 [==============================] - 25s 854us/step - loss: 0.0028 - acc: 0.9992 - val_loss: 0.0078 - val_acc: 0.9979
Epoch 9/10
29058/29058 [==============================] - 25s 850us/step - loss: 0.0015 - acc: 0.9996 - val_loss: 0.0111 - val_acc: 0.9979
Epoch 10/10
29058/29058 [==============================] - 26s 880us/step - loss: 0.0031 - acc: 0.9992 - val_loss: 0.0320 - val_acc: 0.9951
c.开始识别 python3 solve_captchas_with_model.py 由于我们没有桌面，所以代码需要微调，删除这两行
cv2.imshow("Output", output)
cv2.waitKey()
重新运行，效果如下
CAPTCHA text is: 5ULH
CAPTCHA text is: E5KP
CAPTCHA text is: EEGB
CAPTCHA text is: X6RC
CAPTCHA text is: 5AD5
CAPTCHA text is: KAQT
CAPTCHA text is: DC3U
CAPTCHA text is: FLGG
CAPTCHA text is: TFJW
CAPTCHA text is: L8X2