Blog

Looking for a marketer to join the team.
Alternatively, a developer — in that case I’ll handle marketing myself.

If you like the project and see its potential, reach out: [email protected]

Still collecting the dataset. Registered users can now see not only the amount of raw data, but also the amount of labeled data.

3×3:
For many types, I’ve collected enough data. Retrained the model again — quality improved.

4×4:
Automatic labeling doesn’t work well. Even with correct-answer validation, the results are poor. Training on this data adds errors and hurts quality.

So I switched to manual labeling. Labeled a small dataset and trained on it. Lots of crosswalks and traffic lights — eyes get tired, but the results are good.

Satisfied with the model quality. Can start promoting the extension. Even without ads, it already has 1000+ daily users, based on Chrome Web Store data.

Not satisfied with the extension size yet. Need more well-labeled data. Then I can switch back to b0 models and reduce the size to ~40–50 MB.

Version 0.2.10.0 with updated models is now published in the Chrome Store: Free AI ReCaptcha Solver by Raptor (Captcha Plugin)

The dataset is still far from the size I need, but it was enough to retrain the multiclass models. Instead of using 16 separate models for 3×3 and 11 models for 4×4, there is now one unified multimodel for each type of CAPTCHA.

Due to limited training data, I had to switch from b0 to the heavier b1 network. Surprisingly, the accuracy actually improved, and the plugin size dropped from 500 MB to 100 MB. Still, the quality isn’t where I want it yet, so I’m holding off on any advertising for now.

You can check the current dataset size after creating an account:
Register here

I haven't really promoted the extension yet, but people have already started using it - enough for a small but helpful dataset to form. This alone gave a boost to improving the dataset labeling quality.

3×3

I’ve learned how to quickly spot model errors. Usually, seeing the same image 3–5 times is enough to automatically correct a wrong prediction. In rare cases, a single encounter is enough.

Funny thing: sometimes reCAPTCHA insists an object exists even when it clearly doesn’t. For example, it often mistakes cyclists wearing red for a fire hydrant.

4×4

This one was tougher. I finally found a reliable way to determine whether the model marked the correct squares - which is tricky because reCAPTCHA allows a small error margin. That means it may accept a certain answer in one case and reject the exact same answer in another if the error is slightly larger.

To analyze these cases properly, I need to see each image 10–20 times and gather a set of different answers. Because of that, I added a feature to the plugin where it gives slightly varied responses when encountering identical 4×4 tasks - this helps diversify the dataset and makes analysis easier.

Install the extension from the Chrome Web Store:
Free AI reCAPTCHA Solver

- Solving reCAPTCHA on EfficientNet-B0
- Separate models for image types: 16 for 3×3 and 11 for 4×4
- Plans: training EfficientNet-B1 and unification of models to reduce the size of the extension. Register on the website to receive notification about the release of a new version and to download it, distribution will be only among registered users

How tired I am of clicking this recaptcha. I want to try to collect a dataset of images, train a neural network to recognize objects and write an extension in JavaScript that will automatically select the required images.