How Shutterstock Is Training Its System to Help You Find Better Photos
Do you ever get frustrated with the options in your image search? Shutterstock wants to help, and is using deep learning to do it.
Shutterstock created its own convolutional neural network. It is essentially a computer system that is trained to recognize images -- there are millions of specific items such as cats, bicycles, the night sky -- and pull up the most relevant photos.
The image service has begun rolling out a series of tools based on this research. With Reverse Image Search, available today, users can upload an image, either from Shutterstock or another source, and the tool will call up images that look like and have a similar feel to the original photo. Visually Similar Search and Discovery, which makes the results for similar photos or video clips more accurate than ever before, will be rolling out in the coming weeks, according to the service.
Shutterstock VP of Engineering Kevin Lester says that these tools are some of the first steps in moving beyond relying on keywords to identify the best images. Instead, the service has turned to something called Computer Vision, which breaks down the key components of a photo numerically, drawing from its pixel data instead of metadata that is pulled from those tags and keywords.
"Our collection is getting very large. If we don't expand our searchable images beyond keywords, we're heading to a point where it's going to be hard for those images to get visibility with customers," Lester says. "There is a ceiling that I think the industry is kind of heading towards, that in order to break through it you have to change the way you think about your media."
Lawrence Lazare, product director for search and discovery, agrees. "As we consume more and more data, findability is huge." He also noted that Computer Vision is making Shutterstock more competitive, since the company won't have to purchase a similar system from another business in the space, like Getty or Corbis Images.
Shutterstock has more than 70 million images and 4 million video clips in its collection, with 60,000 new pieces of content added every day. The company sells 4.7 images per second to its active customer base of 1.4 million people in 150 countries.