crnn-captcha-solver

Solving captcha using CRNN + CTC loss

The model has been trained on 3500 examples which look like this:

They've been binarized to remove background noise:

Augmentations used: random small-angle rotations and piecewise affine transformation - to simulate distortions in captcha images.

The model shows about 76% accuracy on the validation set.

To use:

python inference.py --model_path ./netCRNN_trained.pth --image_path ./image.jpg

Train on your own data

Move your data to train and test dirs. To handle the problem of images with the same texts, your files should be named as <captcha_text>_<index>.png, for example veq4n_2949.png
Run tool/create_dataset.py for train and test and specify directory for output lmdb dataset
You can specify your character set in alphabets.py
You can change learning parameters in params.py

Use inference.py to predict your captcha images. Specify plot_needed parameter to visualize output of the model.

For example, for given image:

we can visualize output probability distributions for channels s, 4, d, 7 and - (empty):