alt Hacker Newsprobably a CLIP. i've used it recently to build a search engine for google street view.
just create an embedding of your drawing, then an embedding of the actual suspect. and check the distance. if it's around 5-10 drawings you know are similar - then it's a match. if not then not. embedings of matched drawings can also be used to adjust how close the drawing has to be to the image.
Wouldn't that work if you already know what to search for? Wouldn't all drawings for streets or overview maps looks very simillar to the point where false positives will be unavoidable?