mercredi 30 mars 2016

5 Things You Didn’t Know About Image Recognition (and Why It Matters for Brands)

As Internet analyst Mary Meeker pointed out a couple of years ago, around 2 billion images are shared online every day on websites, social networks, and chat apps. That’s a lot for brands to keep on top of, especially considering that 80 percent of all images shared online lack identifying hashtags or text. But thanks to recent advances in the artificial intelligence and machine learning space – namely, image recognition – it’s getting easier for brands make the most of the increasingly visual web. To help you understand this increasingly powerful technology, take a look at our brief guide.

Here’s what we mean when we talk about image recognition.

There are varying definitions of image recognition, which can be confusing, but basically it has to do using computer algorithms to “glance” at images and instantly make sense of them. (There are other terms for it, like computer vision, but we’ll stick to image recognition.) Think of an infant learning how to identify people and things—mama, dada, doggy, apple, SpongeBob, Kardashian—and then imagine that learning curve progressing at mind-boggling speed, and at massive scale, and you can begin to grasp the basics of image recognition.

Image recognition is about, at its core, pattern recognition.

Like babies, computers can be taught to recognize people and things by watching for patterns. All the variables that distinguish mommy’s face from daddy’s, for instance—size, shape, skin tone, the distance between eyes, etc.—can be quantified, and then analyzed, by a computer. Right now you’re probably carrying (or reading this on) a smartphone with a built-in camera that can do a rudimentary version of image recognition in the form of very basic facial recognition—which allows the camera to autofocus on faces instead of background or foreground objects in a scene. It can do that by analyzing the pixels—the visual information translated into data—and deciphering patterns: This stuff over here is a probably a human face; that other stuff over there probably isn’t.

The need for image recognition is exploding because visual culture is exploding.

Recommended for YouWebcast: Advanced Facebook Growth Hacking

In its earliest days, the internet was almost entirely text-centric. There’s still, of course, an endless and ever-growing amount of text on the web, but given the rise of smartphones with awesome built-in cameras, and social media channels that make image sharing stupid-easy, millions of people are regularly letting photos do the talking for them in lieu of words. Facebook alone says that users upload more than 350 million photos to its servers each day. In May, British photo-printing service Photoworld did some math and concluded that “It would take you 10 years to view all the photos shared on Snapchat in the last hour”—presuming 10 seconds of viewing time per snap—and “by the time you’d viewed those, another 880,000 years’ worth of photos would have been shared.” Consider the other image-centric platforms in the social space, including Tumblr and Pinterest, and it’s enough to make your brain (and your eyes) hurt.

There’s no way mere humans could ever hope to keep up—which is why computer scientists are teaching machines how to get smarter and smarter about making sense of images.

Most images are completely, utterly irrelevant to you.

That’s a sad fact that’s just getting more and more true as our collective social-media-driven obsession with photos continues to grow. And that’s where image recognition really begins to matter: Once you can start to automate the process of sorting and making sense of images using computer algorithms, then you can start telling computer systems to only show you relevant images.

Today’s image recognition technology looks for more than just faces, and can identify everything from objects (tennis ball, motorcycle) and scenes (mountains, highway, hotel room) to anatomical structure (facial features, body parts) and even logos, with gender, cartoons, and car make and model on the way.

For humans—and brands—relevance is everything.

If you’re a brand, searching the text-centric web to find out what consumers are saying about you is a pretty straightforward process. Sure, you can do image searches on Google and other search engines to find visual references to your brand, but those tend to be images that have been tagged or labeled by a human—e.g., a news site’s post about Starbucks might have a photo of a Starbucks store that’s been captioned by a photojournalist or editor.

But consider a consumer who uploads a photo of her favorite Starbucks drink, complete with a heart drawn on it, while not bothering to caption or hashtag it. That’s a wordless expression of love for a brand that wouldn’t show up in a traditional text-centric search. According to research, 80 percent of images shared online lack basic identifying text or hashtags.

And this is where image recognition can help. GumGum Social’s visual listening technology, for example, can recognize the presence of a Starbucks logo in an image – in this case, a coffee cup — and then automatically tag it and sort it into a virtual pile labeled, say, “All the Starbucks-related photos on social media from the past 72 hours.” Image recognition is also used to locate relevant editorial images where in-image ads can be served. In the case of, say, car brands, that can mean “conquesting,” when an ad for, say, a Chevrolet Suburban, is served within an editorial picture of a Ford Expedition headlining an article about summer road trips.

Chevrolet happens to have a relatively distinctive and recognizable logo that includes a bowtie-shaped emblem; this makes it pretty easy for a computer to recognize, most of the time. Where things get tricky is that this technology has to be able to make sense of objects and logos in space — curving a bit, or where only a part of it is showing — which is where image recognition increasingly begins to intersect with a branch of machine learning known as deep learning, which enables computers to teach themselves to identify more and more of what’s going on in pictures, even when it isn’t obvious. And the way that computers learn this is by analyzing massive amounts of images; millions and millions of them. The good news is that, with 2 billion images shared online every day, there are plenty of pictures that computers can learn from.

It’s totally the stuff of science fiction, but the visual revolution online – not to mention progress in artificial intelligence – is happening for real right now. Progressive brands will make the most of it.

Let's block ads! (Why?)



5 Things You Didn’t Know About Image Recognition (and Why It Matters for Brands)

Aucun commentaire:

Enregistrer un commentaire