Recognition of Text Verification Code Based on Residual Network

In this project, a VGG-Like model, a ResNet-34-Like model and a ResNet-50-Like model are used to break an enhanced captcha from the "captcha" project.

Requirements:

To run this project, you need a computer with Nvidia GPU that support CUDA, and runs the Linux operating system.

PyTorch and Torchvision witb CUDA support is the main library used in this project. Pillow, Numpy and Scipy are also used to generate captcha images. Tensorboard is also used to view and export the result of loss and accuracy of the project.

To install all the libraries except Torchvision on Archlinux or Manjaro, run the following command in terminal:

sudo pacman -Syu python-pytorch-cuda python-scipy python-numpy python-pillow tensorboard

You can install Torchvision via AUR is possible, but to use a Python virtual environment is recommended. To set a blank virtual environment, run the following command:

python -m venv venv --system-site-packages

To activate the virtual environment, we could switch the project directory and use the following command

source ./venv/bin/activate  

Then we could install the Torchvision via PIP command:

pip install torchvision

Run the code

Firstly, you need to switch to the project directory, and then activate the virtual environment via the command:

source ./venv/bin/activate  

Before running the main function, there are small adjustments in the main.py file. The file includes 3 models:

model = CnnModel(CAPTCHA_CHAR_LEN, CAPTCHA_STR_LEN, input_shape=(3, PIC_HEIGHT, PIC_WIDTH))
model = ResNet50Model(CAPTCHA_CHAR_LEN,CAPTCHA_STR_LEN, input_shape=(3, PIC_HEIGHT, PIC_WIDTH))
model = ResNet34Model(CAPTCHA_CHAR_LEN,CAPTCHA_STR_LEN, input_shape=(3, PIC_HEIGHT, PIC_WIDTH))

You need to choose one model and comment out other two. Then we need to change where to store the training results, we need to change the following variable to the directory to store the results:

STORE_DIRECTORY = "log/model_1/"

Then we could start up the main project via the following command:

python main.py

It takes some hours to run the programme, then we could run the command to launch tensorboard and view the result:

tendorboard --logdir ./