Google turns to machine studying to advance translation of textual content out in the true world • TechCrunch

Google is giving its translation service an improve with a brand new machine learning-powered addition that can permit customers to extra simply translate textual content that seems in the true world, like on storefronts, menus, paperwork, enterprise playing cards and different gadgets. As a substitute of protecting up the unique textual content with the interpretation — the brand new characteristic will as a substitute well overly the translated textual content on high of the picture, whereas additionally rebuilding the pixels beneath with an A.I.-generated background to make the method of studying the interpretation really feel extra pure.

“Usually it’s that mixture of the phrase plus the context — just like the background picture — that basically brings that means to what you’re seeing,” defined Cathy Edwards, VP and GM of Google Search, in a briefing forward of as we speak’s announcement. “You don’t need to translate a textual content to cowl up that essential context that may come by within the photos,” she mentioned.

To make this course of work, Google is utilizing a machine studying expertise often called generative adversarial networks, in any other case often called GAN fashions — the identical expertise that powers the Magic Eraser characteristic to take away objects from images taken on the Google Pixel smartphones. This development will permit Google to now mix the translated textual content into even very advanced photos, making the interpretation really feel pure and seamless, the corporate says. It ought to appear as in the event you’re wanting on the merchandise or object itself with translated textual content, not an overlay obscuring the picture.

The characteristic is one other growth that appears to level to Google’s plans to additional put money into the creation of recent AR glasses, as a capability to translate textual content in the true world may very well be a key promoting level for such a tool. The corporate famous that each month, folks use Google to translate textual content and pictures over a billion occasions in additional than 100 languages. It additionally this 12 months started testing AR prototypes in public settings with a handful of workers and trusted testers, it mentioned.

Whereas there’s apparent demand for higher translation, it’s not clear if customers will favor to make use of their smartphone for translations reasonably than particular eyewear. In spite of everything, Google’s first entry into the smartglasses house, Google Glass, in the end failed as a client product.

Google didn’t converse to its long-term plans for the interpretation characteristic as we speak, noting solely that it could arrive someday later this 12 months.