Try yolo tiny. Must detect well if it is trained properly. But do you think you need a nn model? Scanners read barcodes without neural networks.

But scanners do not read digital data, or even find out the position of the object, and they use light reflection/absortion, there is no digital information reading.

Quick discovery: https://note.nkmk.me/en/python-opencv-barcode/

I would say, if it is not a very wild environment, nn may be excessive.

I need to run it in the browser, I thought it was clear from the post. I found a JS port but it seems flacky. They use CNN for 2d detection though, at least in some cases.

How could we recognize that you want whole processing to be done purely in-browser? Anyway, just google for "webassembly barcode reader". There are implementations available.