This repository hosts our in-house trained captcha model, using over 80,000 captcha images gathered from bots.
The model is only 13.2 MB (12.6 MiB) in size, which is only 25% of the size of ddddocr's common.onnx model, which is 54.1 MB (51.58 MiB).
When testing the model using a dataset of 950 labeled captcha images (not used in training), it achieved an accuracy of 81.79% (777/950).
For comparison, ddddocr only managed to achieve an accuracy of 69.68% (662/950).
Compared to ddddocr, our model takes 41.7ms1 to solve a captcha, which is 28.3ms slower than ddddocr's 13.4ms2 per captcha.
Although slower, that difference is negligible and will not be noticable.
View the example on how to use the model.