Computers That Can Recognize Sketches Faster Than Humans


Quickly interpreting and recognizing an image, once thought a very human capability, is now being done by computer software — and faster and better than we can manage it, researchers say.

Computer scientists at Queen Mary University of London say they've developed an artificial neural network they've dubbed "Sketch-a-Net" that can correctly identify the subjects of simple hand-drawn sketches 74.9 percent of the time, just edging out a human sketch recognition benchmark of 73.1 percent.

Sketch-a-Net could go beyond simply identifying a sketch as a bird, the researchers report in their study, and could recognize the difference between a standing bird, a flying bird, a seagull and a pigeon.

They did that with 42.5 percent accuracy versus humans, who could only rack up a 24.8 percent "spot the kind of bird" success rate.

The technology could have applications in commerce, criminal investigations or even finding the particular cat picture you remember from the Internet, the researchers say.

They trained their neural network by showing it thousands of sketches in a variety of styles and with large numbers of subjects, which were matched to keywords.

Sketching could become more prevalent than keyword searches as more and more people interact with devices that utilize touchscreens, they suggest; people looking for a type of furniture or fashion accessory could simply sketch it to launch a search.

Sketch-a-Net is what is termed a "deep neural network," designed to emulate as closely as possible the processing techniques of the human brain.

It uses eight successive layers of processing activity to break a sketch down into component parts, such as the order in which the line strokes were drawn, to categorize and recognize its subject, the researchers explain.

Sketches are challenging to identify because they can vary depending on the person drawing them, they are somewhat abstract, and they consist of simple black or white lines rather than pixels seen in a photo.

"It's exciting that our computer program can solve the task even better than humans can," says study co-author Timothy Hospedales of the University's School of Electronic Engineering and Computer Science.

"Sketches are an interesting area to study because they have been used since prehistoric times for communication and now, with the increase in use of touchscreens, they are becoming a much more common communication tool again."

ⓒ 2018 All rights reserved. Do not reproduce without permission.
Real Time Analytics