Living in the past...
Most Usenet binary content really is in binary, not text.
The old text-based methods of encoding binary fell out of use years ago - most of it is now yEncoded, which uses virtually all of the 256 possible values of each byte. Therefore it would be trivially easy to filter out. However, once you start filtering out yEncode, people will simply stop using it and go back to Uuencode or Base64 which are pure printable text encodings. But even then, it's trivially easy to spot encoded binary in these forms. So, people will develop encodings that cannot be trivially detected - they'll take up rather more storage space and bandwidth, but that's not exactly expensive these days.
The other thing about binary content is that it's BIG. Yes of course it's split into multiple small messages, but even then they have to be reassembled somehow and so the message subjects contain this information. This makes the typical "infringing content" trivially easy to detect simply by its size. Of course, it would be possible to invent a scheme where the message subjects were assigned at random. You'd still need indexing sites though, which could then be the subject of prosecutions.