Nestert/pytorch-captcha-recognition

Generate captcha-label pairs automatically and train ResNet model to recognize captcha (accuracy: 90%)

Captcha recognition with CNN (ResNet)

Dependence:

• System: windows 10
• GPU: RTX 2080
• Python: 3.6.5
• pytorch: 1.0.0-gpu
• opencv
• captcha

General idea:

• Generate captcha-label pairs with captcha module
• Train CNN (ResNet) model on generative training dataset
• Main reference: CNN_captcha

Data preparation/processing

python generate_captcha.py
- custom_captcha_image.py
- parameters.py

Instead of importing captcha.image in captcha module directly, new custom_captcha_image.py is setup (based on image.py)

• Captcha with target web styles could be generated easily by customizing some parameters
  • image size
  • font types and size
  • noisy line
  • noisy dots
• Avoid revising image.py (sys file) every time
  • location of image.py: Path to Anaconda3\Lib\site-packages\captcha

200K captcha images are generated for training

10K captcha images are generated for test (validation)

Modelling and summary

python -m visdom.server  # run in cmd, avoid errors reported
python main.py
- model.py
- train.py
- parameters.py
- dataset.py
- Visdom.py

Models:

• setup ResNet model as main structure
• use softmax for multi-class classification (single-char level)
• treat 4 chars (one captcha content) as one example for accuracy calculation (uppercase and lowercase is recognized different)
• size of saved model is around 40 M
• test accuarcy on test dataset reaches 0.75 after 5 epochs

python userTest.py

Results:

• test accuarcy on golden_set (labelled manually on real captcha images):
  • 0.30, uppercase and lowercase is recognized different
  • 0.94, convert to lowercase both (it is the real application scenarios)

Online api

python userPred.py

Some changes made for runtime prediction compared to userTest.py

• use test data directly rather than dataLoader
• set num_workers=0 in dataLoader when used in prediction: In windows, DataLoader with num_workers > 0 is extremely slow (50 times slower)

python api_demo.py

Implement with flask module and single response time is around 60 ms (test with Postman)

More thinking

Why is it that current classfication result is much better than python-captcha-recognition?

• Labels are always correct in custom_captcha_image rather than ocr methods
• Uppercase and lowercase is recognized different in current training process (total 62 classes) while all the letters are converted to lowercase and then do classification (total 36 classes) before
• Num of training examples is much higher and deep learning model (ResNet here) is more suitable
• End-to-end training: raw captcha iamge -> labels, no image preprocessing involved
• It is scalable to different web styles (changes in captcha generation is required only)