trekhleb / links-detector Goto Github PK

Currently the HTTPS-detection model is trained just on one font. We need to add a support for more fonts to make the model more generalizable.

Hello!
Thanks for the whole article. I would like to try it.... Would it be possible to get the dataset?

Best
Holger

For now Links Detector recognises only the links that fit into one line. For example:

...some random text https://test.com/sub/path?param=true some other random text...

In the books, however, you may see the links that are wrapped into two lines like this:

...some random text https://test.com/sub/
path?param=true some other random text...

It would be nice if Links Detector will recognise multiline links as well.

The application size is too big right now. The first load may take up to 20Mb which may be too problematic for slow networks.

Currently it may take close to 20Mb of a traffic to do the initial load of the application. By making the app to be PWA (progressive web app) friendly it is possible to load the up many time faster on consecutive loads.

Currently, depending on the device, the recognition performance is close to 1-2 frames per second (fps) which feels really slow. It would be good to at least double it (up to 4+ fps).

Currently for the app to work the text needs to be black on a white background. However it is a common case to see light links on dark backgrounds. The model dataset needs to be extended to support that.

Currently the model is trained on the links that are not underlined. Model dataset needs to be extended to support underline links detection as well.

For now Links Detector is trained to recognise the links that start from https://. The model needs to be re-trained to support http:// links as well.

For now, for performance reasons and limitations, Links Detector supports simultaneous recognition of up to two links on the page. This number is hardcoded and may be increased easily but it will reduce the app performance (which is pretty slow already). However It would be nice to improve the OCR performance to allow more than 2 simultaneous recognitions per page.