#Identification codes#
Solution: verification code into individual character training
Encountered a problem: Code character sizes or overlap
To solve the above problem: by CNN (convolution neural network) is the direct end of identification undivided
Processing Code: The image binarization
Enter the code of information: image size, the largest letter, representation
Selected from the 2 * 2 (three layers and a convolution layer fully connected) and the maximum cell layer dropout layer selected convolution kernel size 5 * 5
About softmax and sigmoid:
Use softmax final normalization operation, namely "Let the great value of classification possibilities becomes larger, so that small possibility of component values become smaller."
Some common file name:
cfg configuration file
cnn_sys cnn network structure
iterative dataset can data_iter
gen_captcha verification code generator
predict loaded trained model
train the model train
utils Public Use