What is the capthca

Captcha is image generated by chars randomly and proper voice to avoid the attack from the robot.

The picture above is a common captcha,including four digital numbers.
#How to recognize the captcha There three main methods:

• find the loopholes of the captcha
• spilt the captcha into single char, and then recognize the chars
• regard the captha as the whole, and recognzie it directly

Tesseract OCR and OpenCV use the second method. But the captcha is more and more complex now. There is a common phenomenon that the chars in the captcha interlace with one another, so the second method gets a low accuracy. This program focuses on the third method.

Generate the captcha to train

Acquire massive capthca by hunman is unrealistic. Fortunately, we can easily generate the captcha by the python package captcha.See the get_train_set.py in details.

Build the cnn model

I build the network with keras, using the theano backend. I think use keras raher than theano to code is much more convenient.
The network includes three convolution layers and two full-connected layers. Owing to CNN needs a large amount of samples to train, with the limitation of time and resources, I only use digital number as the char set of the captcha. So output of the model has 410 numbers, while is will have 462 numbers if the char set includes ppercase and lowercase letters.

# 3 conv layer
model.add(Conv2D(nb_filters1, (kernel_size[0], kernel_size[1]), padding='valid', input_shape=input_shape))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=pool_size))
model.add(Dropout(0.25))

model.add(Conv2D(nb_filters2, (kernel_size[0], kernel_size[1])))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=pool_size))
model.add(Dropout(0.25))

model.add(Conv2D(nb_filters3, (kernel_size[0], kernel_size[1])))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=pool_size))
model.add(Dropout(0.25))

# Fully connected layer
model.add(Flatten())
model.add(Dense(1024*MAX_CAPTCHA))
model.add(Dense(512*MAX_CAPTCHA))
model.add(Activation('relu'))
model.add(Dropout(0.25))
model.add(Dense(MAX_CAPTCHA*CHAR_SET_LEN))
model.add(Activation('softmax'))

model.compile(loss='categorical_crossentropy',
              optimizer='adadelta',
              metrics=['accuracy'])

The training result

I use 18000 training samples and 6000 test samples. After 64 epochs of training, the model obtains the correct rate of 97.91% on the test data.There is part of the result following:  

We can see the program successfully recognize "4063" , but fail "7229". The chat '7' and '2' are similar, maybe we need more data to train.

Try yourself

Environment

My Environment is Windows with Anaconda. Anaconda should install the package theano,keras and capthca.

Run my program

• get_captcha.py: Generate the sample captchas
• captcha_train.py: Train the model
• eval_test.py: Solve an image

	python eval_test.py data/50_7136.png

It works pretty well in order to solve captcha created from other libraries such phpcaptcha or using letters and more than 4 characters :)