Awesome

CAPTCHA

Newbie about Deep Learning and TensorFlow?

Boring with MNIST?

Want a more interesting and complicated application?

This is for you. This repo contains a cnn model for recognizing numbers of captcha

WHAT IS CAPTCHA

CAPTCHA is kind of images that contains chars and digits for people to recognize, it is used in website log in to test you whether you are a robot or a person. In this repo we will develop a small convolutional neural network with TensorFlow to recognize it.

For simplicity, images will only contain four digits with noise

we say a image is classified correctly if and only if four digits inside this image are all classified correctly

Two sample images are listed below

requirements

python 2.7 with following packages installed should work fine

1. numpy
2. TensorFlow(verison >= 1.4) (because we will use tf.data)
3. captcha(you can install it with pip install captcha)

(anaconda environment is strongly recommended for managing these packages)

windows and python 3.X are not tested but should be OK.

GPU is not a must, but without it, training might be very slow.

SOME FEATURES

• Train and validation images are generated on the fly, doesn't need to download any big datasets.

• Inputs of model is built on top of tf.data instead of old queue-based api, so reading the code combined with the official document about tf.data will help you understand how to write it yourself.

• Very short code, easy to read.

USAGE

First clone this repo

git clone https://github.com/zakizhou/CAPTCHA

Before run training, training and validation images should be generated, change to the root dir of this repo and run

cd CAPTCHA
mkdir -p images/train
mkdir -p images/validation
mkdir -p tfrecords
mkdir -p save
python captcha_producer.py -n 30000 -p images/train

This will generate 30000 training images in the images/train/ and also convert infomation about these images into tfrecords/train.tfrecords file.

for validation set:

python captcha_producer.py -n 3000 -p "images/validation"

Now you can run this model with

python captcha_train.py

Result

After 10000 steps (you can manually change num of steps in captcha_train.py file) training on single GTX1060, this model achieved around 70% accuracy, adjusting the scale of parameters or adding dropout should still improve this performance

Details about files

• images/train and images/validation will contains generated train and validation images
• tfrecords will contains about generated tfrecords(train.tfrecords, validation.tfrecords)
• save will contains saved model after training
• captcha_producer.py is used to generate images and tfrecords
• captcha_model.py contains utils functions for defining the model
• captcha_data.py is used for build input for model
• captcha_config.py contains configs for model
• captcha_train.py is used for training model

TODO

• add multi gpu training code
• add tensorboard code
• add functions for keep on training after shutdown