Crack Captcha

The captcha we about to crack is very simple, for example:

Model Accuracy History

Commit Hash	Accuracy	Remark
f15750	50.5208%	Batch size 64, 20k.
f15750	55.4688%	Batch size 128, 20k.
not save	48.4375%	Disable random translate, 20k.
bf6b08	45.3125%	Use VGG16, 10k.

The loss don't drop, always ~2.7. :(

TODOs

• Construct the model.
• Crawl the captcha dataset.
• Label the dataset.
• Make it runnable.
• Finish the model training.
• Construct API for the trained model.
• Data augmentation: add salt & pepper noise.

Others

I referred xmcp/elective-dataset-2021spring when crafting this project.