Google’s Pixel phone has one hell of a camera, and one of the reasons for this is AI. Google has used its machine learning talent to squeeze better images out of a tiny smartphone lens, including its portrait mode shots, with blurred backgrounds and pin-sharp subjects
Now, Google has open-sourced a lump of code named DeepLab-v3+ that it says will help others recreate the same effect. (Although, this is not the same tech that Google itself uses in the Pixel phones — see the correction note at the bottom of the article.) DeepLab-v3+ is an image segmentation tool built using convolutional neural networks, or CNNs: a machine learning method that’s particularly good at analyzing visual data. Image segmentation analyzes objects within a picture, and splits them apart; dividing foreground elements from background elements. This can then be used to create ‘bokeh’ style photographs
As Google software engineers Liang-Chieh Chen and Yukun Zhu explain, image segmentation has improved rapidly with the recent deep-learning boom, reaching “accuracy levels that were hard to imagine even five years [ago].” The company says it hopes that by publicly sharing the system “other groups in academia and industry [will be able] to reproduce and further improve” on Google’s work.
At the very least, opening up this piece of software to the community should help app developers who need some lickety-split image segmentation, just like Google does it.