Photo by HP
A powerful portfolio of HP Labs-generated computer vision algorithms is now being deployed at scale by major HP customers. At the same time, the HP Labs team behind the technology is taking the research underpinning it in new directions.
The imaging portfolio, known as HP Pixel Intelligence enables capabilities such as accurately locating, analyzing, and grouping faces and objects, and automated image layout and cropping. HP itself has used many of these algorithms in its own imaging products. But now other companies are adding these capabilities to their own products and services.
One of the largest adoptions to date is by CEWE, Europe's leading photo service provider and partner for online printing. The company is deploying HP Pixel Intelligence to offer its customers enhanced facial recognition in photographs they are looking to print.
“CEWE is the largest photo service provider in Europe, printing millions of photobooks and calendars a year while enabling compelling storytelling for our customers,” says CEWE CTO Dr. Reiner Fageth. “Therefore, we see AI as an opportunity to enable intelligent organization of photos, and make it easier for our customers to create photo products. We are very pleased with our partnership with HP on AI. HP Pixel Intelligence can accurately group faces, and can be integrated with our products without cloud API calls.”
“HP today has installations of Pixel Intelligence by commercial print customers across several geographic regions, helping users with optimized and automated photobook creation,” adds Dr. Qian Lin, Distinguished Technologist for Computer Vision and Deep Learning Research in HP’s Artificial Intelligence and Emerging Compute Lab.
In an age of digital abundance, printed photobooks have taken on a new significance for many families, suggests Tony Lewis, Global Head of the Artificial Intelligence & Emerging Compute Lab.
“The technologies developed by Dr. Lin and her team are helping our customers organize, preserve, and share their most precious memories with their close friends and family members, which is one of the most important aspects of the human experience,” he says.
Lin’s team continue to develop the algorithms underlying the Pixel Intelligence portfolio and she’s hopeful that new and improved algorithms will soon find their way into HP’s next generation of consumer printing products.
“Because HP does so much work with imaging, we have a huge opportunity to use artificial intelligence (AI) and deep learning to offer our customers better experiences,” Lin says.
Key to that effort is finding ways to run the compute-intense algorithms that enable these experiences on devices like phones, laptops, PCs, and tablets. Most AI-driven photo analysis is processed on a centralized cloud server. That requires back-and-forth communication that slows down the process considerably. The HP Labs algorithms, in contrast, are optimized to work with both precision and high speed on relatively low-powered “edge” devices.
“When we demo our work people ask how come our algorithms can connect back to the cloud so quickly,” Lin reports. “I tell them we’re running these very complicated artificial intelligence technologies just on local devices – that’s why it runs so fast.”
The HP Labs group is also looking to expand their application of deep learning and artificial intelligence to other core printing tasks. One promising area of research is helping better identify printing defects. These often show up as lines, spots, or other artifacts in a printed image or document. The challenge is to distinguish lines and spots that were intended to be on the page from those caused by a defect that needs to be addressed.
“We’ve done this by deploying a “semantic segmentation” algorithm that we train to identify the different objects in a 2D image,” explains Lin. “So it might “see” four people, a wall, and some bookshelves and therefore won’t flag the lines of the wall or the shelves as problems, but it will highlight a line that goes “through” any of the people in the image.”
The algorithm creates “print defect maps” that can be used in a variety of applications, from identification of defective parts to print quality optimization. Lin’s team is collaborating with a community of print quality engineers across the company on these applications.