New capcha model just dropped!

[JonseyJones]

It has a big black circle with a random position.

[sfeed1095]

you vil align ze circle and you vil like it

[CaptainChicky]

well if someone is able to manually compile a dataset of captchas+solutions (rip) they could retrain the nn

[LittleEndu]

how many examples would be required to retrain the model? Current model even gets the circle correct some of the time

[moffatman]

16k images of new captcha, it was able to converge since since 4k+, surprisingly easy https://captcha.chance.surf/bundle_16kblack/images.zip https://captcha.chance.surf/bundle_16kblack/model.h5

[drunohazarb]

@moffatman Thanks a bunch for training the model!

7826d5ae53da72a6cf64dd2a82a4a3f5aec557b9

[Yukariin]

it was able to converge since since 4k+, surprisingly easy

Was it really? Accuracy for old captchas (still in use) is ~78% but feels even lower - like 3-4 of 10 captchas being solved correctly. Fine-tuned @moffatman 16k model on combined dataset - 10k old + 16k new captchas from @moffatman and 3.5k old captchas from @coomdev. Achieves 98.7% (7388/7485) accuracy. model notebook

Also I'd recommend to use latest trained model to sanitise your own dataset - both old 10k and new 16k (and coomtech's one) contain some misaligned and/or mislabeled captchas.

[moffatman]

@Yukariin Ah, a hybrid model, really interesting! I didn't realize the simpler captcha still gets served sometimes. I have a lot more data, probably a million+ old captchas, and growing number of new captchas (50k atm). So I will play around with it. Thanks for all your contribution here!!

I did notice some misaligned, since I get most of these from kuroba users, and it doesn't use optimal alignment method.