Captcha Recgnition

A fast and easy way to recognize captcha image, the cnn-model only has 5 layers, with an accuracy of 90%.

• End to end, all you need is one neural network
• Low delay, 1ms for one image
• Lite model, only 5 layers
• Better accuracy, 90%
• Easy to use, just two line

Model provided

• Support 4 chars captcha, for example

• 

• model structure

• 

Dependencies

• tensorflow2
• numpy
• Pillow

Quick Start

1. install all the package below
2. then

   import Decaptcha
   
   # for single image:
   # img_path = './captcha.png'
   y = Decaptcha.dec_batch(img_path)
   print("the value of image {} is {}".format(img_path,y))
   
   # for batch images:
   # img_src = './sample/'
   Y = DeCaptcha.dec_batch(img_src)
   for y in Y:
       print(y)
   
   

Get Dataset

• Here, I use captcha, a pypi package to generate specific captcha images.
• To get your own dataset, just use the function gen_imgs from DataUtils.py

gen_imgs(src="./sample/", count=40000, width=100, height=60, image_suffix="png")

About model training

• Custom loss function
  • To get the model better and more quickly fit the captcha data, I have customed the CE loss function,as follow:

    def captcha_loss(y,pred):
        y_pred = tf.reshape(pred, [-1, max_chars, len(char_set)]) 
        y_true = tf.reshape(y, [-1, max_chars, len(char_set)]) 
        loss = tf.keras.losses.binary_crossentropy(y_true,y_pred,label_smoothing=0.1)   #(BATCH_SIZE, max_chars,)
        loss = tf.reduce_sum(loss,axis=1)      #(BATCH_SIZE,)
        loss = tf.reduce_mean(loss)        #()
        return  loss
    

• Label smoothing
  • set 0.075~0.25. a trick which advances the valid_acc 10 more points.

• valid accuracy plot
  • 

Development

• First version, 2020.6.27