Proctor & Gamble scientist looks to the future of machine vision
Boston, MA. Where will vision technology be in two to seven years? That’s a key concern of Steve Varga, principal scientist, Imaging and Instrumentation R&D, Procter & Gamble. Delivering a keynote address Tuesday morning at The Vision Show, he said he examines the industry’s technology roadmap, and if there are gaps, he establishes partnerships to develop the technologies he needs to gain a competitive advantage.
He said the vision industry today is in the position of servo-motor technology 20 years ago, when servos were replacing mechanical drives. Prospective customers for servo systems said they are great, but they are too expensive, too complicated (requiring experts), and not completely trusted—they are not reliable enough to run unsupervised.
He described a vision-system application today: looking for wrinkles in a label. The system can look for a dark-to-light transition, for example, to determine the edge of the label, but it has no understanding of the concept of wrinkling. It must be programmed to determine standard deviations of intensity that would indicate a wrinkle. Tomorrow, he said, imaging systems will be able to “perceive” and “understand” what they see through big data and deep learning. The advances will be spurred on by vision toolkits from companies like National Instruments as well as FPGAs and dedicated GPUs.
With evolving technology, he said, machines will tackle such problems as figure-ground segmentation and identifying amodal lines—lines that don’t exist in an image but that may be inferred from the arrangement of other image elements. Similarly, he said, machines will learn to perceive color, not just wavelengths of light. “Light is real, color is not,” he said. He described the work of Dr. Beau Lotto in teaching computers how to learn color.
He cited the example of the Craik-O’brien Cornsweet effect and boundary transients. The brain, he said, is good at filling in features we don’t sense (an occluded red square on a checker board, for example) and is used to things being continuous.
He traced AI back to the work of Arthur Samuel in 1949, describing Samuel’s success in teaching computers how to play chess. The technology has advanced to IMB Watson, and from a machine-vision perspective, computers can recognize images of teddy bears, blankets, and girls. Computers can even infer motion—for example, from a still image of a woman dancing in a field of flowers.
He asked, “What’s next? Deep learning can seem to learn about just about anything.” He said Proctor & Gamble is saving 4 million images per second worldwide, resulting in a database that will help machines learn to do things we don’t even know how to code.