3E-Solver-IJCAI

This reposity is official implementation of "3E-Slover: An Effortless, Easy-to-Update, and End-to-End Solver with Semi-Supervised Learning for Breaking Text-Based Captchas" (IJCAI-2022).

Dataset

https://drive.google.com/file/d/1iL4me0ispkHiNsie6Fr_HrnfRh7PsK8y/view?usp=sharing

unzip dataset.zip
mkdir result

Dependency

torch=1.9.0+cu111
torchvision=0.10.0+cu111
numpy=1.21.2
PIL=8.3.1
matplotlib=3.4.3

Code

python FixMatch_2.py --dataset apple --label 500.txt --unlabeled-number 5000 --epoch 400 --lr 0.01
python FixMatch_2.py --dataset ganji-1 --label 500.txt --unlabeled-number 5000 --epoch 400 --lr 0.01
python FixMatch_2.py --dataset google --label 500.txt --unlabeled-number 5000
python FixMatch_2.py --dataset microsoft --label 500.txt --unlabeled-number 5000
python FixMatch_2.py --dataset sina --label 500.txt --unlabeled-number 5000 --epoch 400 --lr 0.01
python FixMatch_2.py --dataset weibo --label 500.txt --unlabeled-number 5000 --epoch 400 --lr 0.01
python FixMatch_2.py --dataset wikipedia --label 500.txt --unlabeled-number 5000 --epoch 400 --lr 0.01
python FixMatch_2.py --dataset yandex --label 500.txt --unlabeled-number 5000 --epoch 400 --lr 0.01

Result

Scheme	Accuracy
Google	76.4%
Microsoft	96.4%
Yandex	90.4%
Wikiepdia	98.5%
Weibo	92.5%
Sina	97.1%
Apple	92.9%
Ganji	99.4%