Techniques for detecting and identifying signs in images

A method for detection and identification, comprising:

• receiving, by software executing on one or more computing devices, an image that includes a sign;
• predicting, by a trained machine learning (ML) detection model of the software, a position in the image and a meta-class of the sign, wherein the meta-class predicted by the trained ML detection model is a class of objects having similar appearance but not all sharing a same Manual on Uniform Traffic Control Devices (MUTCD) code;
• subsequent to the predicting, using one or more modules of the software that are separate from the trained machine learning (ML) detection model to:
• extract text from the region of the image at the position predicted for the sign,
• compare the extracted text to reference text of only a subset of signs of a plurality of signs in a reference database, wherein the subset of signs of are those signs that share the same meta-class as predicted for the sign, to produce a textual prediction of a specific technical name of the sign,
• extract image descriptors of graphics from the region of the image at the position predicted for the sign, and
• compare the extracted image descriptors to reference image descriptors of only the subset of signs in the reference database that share the same meta-class as predicted for the sign, to produce a visual prediction of the specific technical name of the sign;
• combining, by the software, the textual prediction of the specific technical name of the sign and the visual prediction of the specific technical name of the sign to produce a final prediction of the specific technical name of the sign, wherein the specific technical name produced by combining the textual prediction and the visual prediction comprises a MUTCD code;
• and
• outputting, by the software, the position in the image and the specific technical name of the sign.