HUSTCaptchaResolver

Captcha resolver for HUST (Hanoi University of Science and Technology).

Target: to resolve captcha images on 2 websites

• Website ctt-sis
• Website dk-sis

Below is the basic instruction throughout the project.

1. Configuration

• [Optional] Create and activate environments

# Create virtual environment
python -m venv <env-name>

# Activate environment
source <env-name>/bin/activate  # With Linux
.\<env-name>\Scripts\activate   # With Windows

• Install torch with cuda

pip install torch===1.11.0+cu115 torchvision===0.12.0 torchaudio===0.11.0 -f https://download.pytorch.org/whl/torch_stable.html

• Install requirements

pip install -r requirements.txt

2. Tools

• Crawl: Crawl datasets and raw labels. Dataset will be crawled from 2 websites ctt-sis and dk-sis. The amount and sources can be configured in tools/configs/crawl.yml.

bash scripts/crawl.sh

• Relabel: The labels are quite raw. Use the following pre-built Streamlit app to relabel data. App configuration is in tools/configs/relabel.yml

bash scripts/relabel.sh

• Split: Split dataset into train and valid set

bash scripts/split.sh

3. Train model

• Create dataset folder and put dataset inside. You can use your own dataset or use our pre-built dataset (Google Drive)
• Train model. Training configuration is at configs/configs.yml.

bash scripts/train.sh

4. Inference

• Pretrained model (Google Drive) performance:
  • Full sequence accuracy: 99.27%
  • Per character accuracy: 99.88%
• Create directory checkpoints and put the pretrained weights inside
• Run inference

bash scripts/infer.sh

• Request format
  • Default port: 7000 (can be re-configure)
  • API: /
  • Request: POST (form-data)

  "file": <image file> # png, jpg or jpeg
  

  • Response format

  <result> # example: 38275
  

5. Docker

• Rebuild image and run

docker build -t <image-name> .
docker run -p <local-port>:<docker-port>/tcp <image-name>

• Or use my pre-built image

docker run -p <local-port>:7000/tcp theanhtran/hust-captcha-resolver