Tensorflow realizes verification code recognition (I)

effect

Let's see what the finished product looks like first , Finished products , With randomly generated 5W Zhang verification code for training , Training one 40 Multiple epoch, With randomly generated 2W The accuracy of Zhang Zuo's verification code is 98% The appearance of . With simple UI It's like this ：

If you want to complete the code, you can enter the portal ：https://github.com/aolingwen/fuck_verifycode

Generate verification code

To identify the verification code , Of course, the first thing is to get the verification code data ~~, If the verification code data is available, you can write the crawler to catch it , But after catching it, I have to type the verification code myself label, A troubling group . If you don't want to be so troublesome, you can go and find out if there is one ready-made label Good verification code data , Or make it yourself label Good verification code data . Here , I choose to generate it myself . So here is the code generated by the verification code . A simple batch . Just use ImageCaptcha To randomly generate .

# Generate captcha image ,gen_dir It is the directory where the verification code should be landing ,total_size Is the total number of verification codes to be generated ,chars_set Is the character set of the verification code ,chars_num Is the number of characters in a verification code 
def gen_verifycode_img(gen_dir, total_size, chars_set, chars_num, img_height, img_width, font_sizes):
    if not os.path.exists(gen_dir):
        os.makedirs(gen_dir)
    image = ImageCaptcha(width=img_width, height=img_height, font_sizes=font_sizes)
    for i in range(total_size):
        label = ''.join(random.sample(chars_set, chars_num))
        image.write(label, os.path.join(gen_dir, label+'_num'+str(i)+'.png'))

Generate a good verification code , It's like this , Before file name 5 A character is the verification code label：

Now with the verification code , Next, we need to consider how to feed the data to tensorflow 了 , Because of the big data 5W Map , So I used tfrecord Of IO The way , Write next tfrecode How to play .