Two Googlers Teach An AI About Aesthetic Appeal

It can be hard to put words or a value on why a particular image is aesthetically appealing or unappealing, but that's exactly what two Googlers have taught an artificial intelligence program to do. The pair of researchers, who also happen to be college professors, published a white paper about the AI, outlining their goals for it, how it works, and the steps they took to get it to the state that it's currently in. Essentially, the AI program piggybacks off of existing AI conventions in object recognition, and uses a convolutional neural network to aggregate recognition results across groups of objects and bits of scenery throughout a whole image, resulting in an AI program that can effectively rate the quality and aesthetic appeal of an image through repeated training over time, and can even train itself as it's used.

The AI works by identifying each object in a given image to the best of its abilities, and comparing the image itself to known examples of images with a high aesthetic appeal based on a range of metrics. Objects present, color patterns, total saturation, clarity, sharpness, contrast, and a number of other considerations go into the AI's rating of an image. After being fed a specialized data set of base images and tweaks for those images, the AI was able to judge the "best" image from a given batch with a fairly high accuracy rate.

The potential applications for a bot with this sort of ability are actually fairly wide-ranging. An AI with imagery judging capabilities could alert a security crew when a camera feed is starting to degrade, for example, or could provide a completely objective judgment on a piece of art as juxtaposition to human opinions, just for starters. Naturally, this research could build out into a whole new branch of AI-based object recognition. A convolutional neural network was used in this example, but other types of neural networks could be arranged for the same task, including powerful cloud-based nodes with specialized hardware, to produce results on a wider scale, or to learn even more about how humans may judge the value of images.

You May Like These
More Like This:
Android Headlines We Are Hiring Apply Now