卷积神经网络 CNN 验证码破解

使用卷积神经网络CNN的网站验证码破解

项目背景及需求
在对 网站进行访问时，网站会弹出验证码，用来进行核实人机身份，验证码可能会出现在两种场景：

1. 登录网站时
2. 频繁访问时

由于要实现爬取数据，需要对验证码进行识别处理，可以采用以下方式：

• 暂停
• Cookie
• 人工识别
• 程序识别

暂停

优点：实现较简单，通用性好。
缺点：全程需要人工参与，无法实现数据爬取的自动化。

Cookie

优点：实现较简单，通用性好。
缺点：程序启动后，无需人工参与，但仅能处理登录时的验证码，局限较高。

人工识别

优点：实现非常简单，通用性很好。
缺点：需要花费费用，及时性、稳定性等无可控性。

程序识别

优点：无需人工参与，及时性与稳定性等具有可控性。
缺点：实现复杂，通用性差。

为了实现爬取数据的自动化及稳定性，这里使用卷积神经网络CNN来识别验证码程序。

生成验证码

这里使用captcha库模拟生成验证码（在实际项目中要爬取目标网站的验证码进行训练）。

安装

captcha的安装非常简单，执行如下的命令即可：

pip install captcha

captcha需要依赖于pillow库，如果没有安装pillow库，则pillow库会一同安装。

from captcha.image import ImageCaptcha
from PIL import Image

image = ImageCaptcha()
# 返回BytesIO类文件对象。生成参数指定的验证码。
bio = image.generate("abcd")
# 通过Pillow库的Image类的open方法，可以打开类文件对象，返回Pillow中的
# Image对象。
# Image.open(bio)
# 向指定的路径写入验证码文件。第1个参数：验证码的内容。第2个参数：验证码文件的路径。
image.write("abcd", "d:/test.png")

# sring模块的使用
import string
# 返回ascii字符集中所有的小写字母与大写字母。
print(string.ascii_lowercase)
print(string.ascii_uppercase)

abcdefghijklmnopqrstuvwxyz
ABCDEFGHIJKLMNOPQRSTUVWXYZ

执行步骤
首先，生成验证码数据集，包含验证集与测试集。

python datasets/gen_captcha.py -d --npi=4 -n 6

然后，进行训练。

python cnn_n_char.py --data_dir images/char-4-epoch-6/

训练次数
训练10000次，准确率如下：
step 100, training accuracy = 12.50%, testing accuracy = 9.00%
step 200, training accuracy = 10.00%, testing accuracy = 12.50%
step 300, training accuracy = 15.50%, testing accuracy = 9.00%
step 400, training accuracy = 17.50%, testing accuracy = 18.50%
step 500, training accuracy = 20.50%, testing accuracy = 19.75%
step 600, training accuracy = 21.00%, testing accuracy = 23.00%
step 700, training accuracy = 21.00%, testing accuracy = 24.75%
step 800, training accuracy = 29.50%, testing accuracy = 27.00%
step 900, training accuracy = 29.50%, testing accuracy = 32.25%
step 1000, training accuracy = 30.00%, testing accuracy = 30.50%
step 1100, training accuracy = 32.50%, testing accuracy = 31.00%
step 1200, training accuracy = 38.50%, testing accuracy = 38.00%
step 1300, training accuracy = 38.00%, testing accuracy = 33.75%
step 1400, training accuracy = 37.00%, testing accuracy = 36.50%
step 1500, training accuracy = 39.50%, testing accuracy = 33.75%
step 1600, training accuracy = 43.00%, testing accuracy = 34.25%
step 1700, training accuracy = 38.50%, testing accuracy = 37.75%
step 1800, training accuracy = 40.00%, testing accuracy = 38.25%
step 1900, training accuracy = 44.00%, testing accuracy = 40.00%
step 2000, training accuracy = 48.00%, testing accuracy = 41.75%
step 2100, training accuracy = 44.00%, testing accuracy = 46.25%
step 2200, training accuracy = 39.00%, testing accuracy = 46.00%
step 2300, training accuracy = 48.00%, testing accuracy = 45.00%
step 2400, training accuracy = 49.00%, testing accuracy = 45.75%
step 2500, training accuracy = 51.00%, testing accuracy = 47.25%
step 2600, training accuracy = 50.00%, testing accuracy = 47.25%
step 2700, training accuracy = 49.50%, testing accuracy = 44.25%
step 2800, training accuracy = 45.50%, testing accuracy = 51.50%
step 2900, training accuracy = 42.50%, testing accuracy = 41.25%
step 3000, training accuracy = 50.50%, testing accuracy = 47.00%
step 3100, training accuracy = 54.00%, testing accuracy = 49.25%
step 3200, training accuracy = 51.00%, testing accuracy = 48.00%
step 3300, training accuracy = 53.00%, testing accuracy = 47.00%
step 3400, training accuracy = 50.00%, testing accuracy = 52.75%
step 3500, training accuracy = 60.00%, testing accuracy = 49.75%
step 3600, training accuracy = 53.50%, testing accuracy = 46.50%
step 3700, training accuracy = 60.50%, testing accuracy = 53.75%
step 3800, training accuracy = 53.00%, testing accuracy = 54.00%
step 3900, training accuracy = 56.50%, testing accuracy = 48.75%
step 4000, training accuracy = 58.00%, testing accuracy = 50.50%
step 4100, training accuracy = 54.00%, testing accuracy = 49.25%
step 4200, training accuracy = 53.50%, testing accuracy = 46.00%
step 4300, training accuracy = 55.00%, testing accuracy = 56.00%
step 4400, training accuracy = 66.00%, testing accuracy = 52.25%
step 4500, training accuracy = 57.50%, testing accuracy = 60.50%
step 4600, training accuracy = 60.50%, testing accuracy = 54.25%
step 4700, training accuracy = 59.00%, testing accuracy = 58.75%
step 4800, training accuracy = 60.50%, testing accuracy = 57.00%
step 4900, training accuracy = 63.00%, testing accuracy = 57.75%
step 5000, training accuracy = 59.50%, testing accuracy = 61.75%
step 5100, training accuracy = 62.50%, testing accuracy = 55.25%
step 5200, training accuracy = 60.00%, testing accuracy = 60.50%
step 5300, training accuracy = 66.50%, testing accuracy = 57.75%
step 5400, training accuracy = 64.50%, testing accuracy = 53.50%
step 5500, training accuracy = 65.00%, testing accuracy = 56.75%
step 5600, training accuracy = 63.50%, testing accuracy = 51.00%
step 5700, training accuracy = 68.50%, testing accuracy = 62.25%
step 5800, training accuracy = 64.00%, testing accuracy = 57.50%
step 5900, training accuracy = 69.50%, testing accuracy = 59.75%
step 6000, training accuracy = 64.00%, testing accuracy = 54.50%
step 6100, training accuracy = 61.50%, testing accuracy = 62.50%
step 6200, training accuracy = 70.00%, testing accuracy = 60.50%
step 6300, training accuracy = 71.00%, testing accuracy = 54.75%
step 6400, training accuracy = 65.50%, testing accuracy = 57.00%
step 6500, training accuracy = 71.00%, testing accuracy = 62.75%
step 6600, training accuracy = 68.00%, testing accuracy = 64.75%
step 6700, training accuracy = 65.50%, testing accuracy = 59.75%
step 6800, training accuracy = 71.00%, testing accuracy = 61.50%
step 6900, training accuracy = 68.00%, testing accuracy = 62.00%
step 7000, training accuracy = 69.50%, testing accuracy = 60.00%
step 7100, training accuracy = 71.00%, testing accuracy = 60.00%
step 7200, training accuracy = 76.00%, testing accuracy = 62.25%
step 7300, training accuracy = 76.00%, testing accuracy = 66.50%
step 7400, training accuracy = 72.00%, testing accuracy = 62.00%
step 7500, training accuracy = 73.50%, testing accuracy = 64.75%
step 7600, training accuracy = 68.00%, testing accuracy = 62.00%
step 7700, training accuracy = 69.00%, testing accuracy = 63.00%
step 7800, training accuracy = 77.50%, testing accuracy = 66.00%
step 7900, training accuracy = 82.50%, testing accuracy = 65.75%
step 8000, training accuracy = 78.50%, testing accuracy = 66.25%
step 8100, training accuracy = 77.50%, testing accuracy = 66.00%
step 8200, training accuracy = 80.50%, testing accuracy = 58.00%
step 8300, training accuracy = 78.00%, testing accuracy = 66.00%
step 8400, training accuracy = 77.50%, testing accuracy = 65.00%
step 8500, training accuracy = 81.50%, testing accuracy = 63.75%
step 8600, training accuracy = 76.00%, testing accuracy = 63.25%
step 8700, training accuracy = 74.00%, testing accuracy = 63.75%
step 8800, training accuracy = 83.00%, testing accuracy = 60.75%
step 8900, training accuracy = 79.00%, testing accuracy = 62.50%
step 9000, training accuracy = 76.00%, testing accuracy = 63.50%
step 9100, training accuracy = 77.00%, testing accuracy = 66.50%
step 9200, training accuracy = 84.50%, testing accuracy = 64.00%
step 9300, training accuracy = 84.50%, testing accuracy = 68.25%
step 9400, training accuracy = 80.50%, testing accuracy = 67.25%
step 9500, training accuracy = 82.50%, testing accuracy = 69.00%
step 9600, training accuracy = 80.50%, testing accuracy = 65.00%
step 9700, training accuracy = 90.00%, testing accuracy = 70.00%
step 9800, training accuracy = 87.00%, testing accuracy = 63.25%
step 9900, training accuracy = 87.00%, testing accuracy = 65.75%
step 10000, training accuracy = 85.00%, testing accuracy = 68.75%

如果想要训练一个期望的准确率（可能达到的准确率），可以采用无限循环，然后在准确率达到期望值时停止训练。