• Facebook

    500 K / likes

  • Twitter

    1 M / followers

How Much of a Photo Can Be Deleted Before AI Can’t Recognize It?


In a new project that mixes science and art, artistic duo Shinseungback Kimyonghun has created a series of images that have pixels removed until an AI program can no longer recognize the subject — in this case mountains. Impressively, much of the image can be deleted before this happens.

“Shinseungback Kimyonghun” is a Seoul-based artistic duo consisting of engineer Shin Seung Back and artist Kim Yong Hun. The two have many projects that mix AI visual recognition with photography. In one from 2012 called Cloud Face, the two had an AI look at moving clouds in an attempt to pull frames that looked like human faces. In another project called Flower, the AI was shown a series of distorted flower images which could still be recognized by the AI.

In a similar vein to Flower, this latest project, titled Mou ta n, examines the limits of AI’s current object recognition capabilities.

“An unfinished painting by Paul Cézanne, Still Life with Water Jug inspired the project,” Yong Hun tells PetaPixel. “Although almost half of the painting is unpainted, we can still see the objects. Would it be visible to AI as well? If so, how much should it be erased to be unrecognizable to AI? What would it look like to humans then? These questions arose and we applied the idea to mountain images.”

Yong Hun says that he and Back used three different object recognition systems for this project: Google Cloud Vision, detectron2 by Facebook, and Microsoft Azure Computer Vision. When none of the three AI systems can see the mountains in the images, the duo considered the piece complete.

Looking at the mostly-deleted images, it is rather impressive that the AI systems were able to see these as mountains for so long: some of these are nearly indistinguishable even to the human eye. Yong Hun and Beck agree.

“We had to erase more than we expected,” Yong Hun says.

For the team, finding a balance to where the AI wouldn’t be able to recognize it but a human still would was part of the challenge.

“We wanted the erased mountains to still be visible to humans. This boundary was difficult to find. If we erase only a little, the AI sees the mountain. If we erase a lot, humans cannot see the mountain either,” Yong Hun said.

The portions they chose to erase had to be picked in such a way that humans could still recognize them. But Yong Hun says that if they wanted to, they could have deleted so much of the photo that a human wouldn’t be able to see what it was, but an AI could still recognize it.

“The images were erased in the way that humans (at least myself) can see the mountains, but it could have been erased for AI can see but humans can’t,” he says.

The duo believes that as AI gets smarter, it will be harder and harder to create images that humans can still recognize but AI cannot. Part of what they found in this process was that in many cases, AI showed that it may be better at recognition of objects than humans are, and this is especially the case as AI gets smarter.

While certainly interesting as art, this project does point to the power of artificial intelligence. With image recognition and facial recognition getting particularly powerful, it will be part of human society’s choices in the near future to determine the limits of how the technology can be deployed, what it can be used on, and how governments should be expected to use or not use it.

To see Shinseungback Kimyonghun’s full project, check it out here.

Image credits: Photos by Shinseungback Kimyonghun and used with permission.