CNN_PyTorch

Looking for training using Keras? switch to branch master.

Introduction

Convolutional Neural Network using PyTorch for CAPTCHA recognition.

Recognition for this type of CAPTCHA (generated from repo CAPTCHA_generator):

In branch master, we must firstly split the whole image into single letters, and then use letters as training data. So, for different CAPTCHA, we have to find a specific method to split it. The benefits is that while splitting the picture, we also expand the training data samples and reduce the problem size.

Method used in this branch is that we see the whole picture as one training sample with four labels, so it becomes a multilabel-classification problems. See model graph below for detail info.

Achievement

Using 4.8k training set:

Using 10k training set:

Using 20k training set:

Model Graph

Sorry for ugly painting, waiting for your pull requests to make it better :)

Others

• If you have any questions, feel free to post a issue.