Your eyes may fool you. But for a picture to be taken perfectly, there are a lot of things involved. You need to deal with the lights, weather condition, speed and others to get the best shot. You also need to edit them to get the best results out there.
Not just your megapixel smartphones can take pictures equally as good, but to please the eyes, there are indeed some skills required. And results speak for themselves.
Google has long been into Artificial Intelligence. The tech titan has come closer to solving the last puzzle: cropping photos and editing them to get the best picture balance.
Using Machine Perception, the researchers have trained a deep-learning system to identify ~40,000 panoramas photos taken by Google Street View in areas like the Alps, Banff and Jasper National Parks in Canada, Big Sur in California and Yellowstone National Park.
The AI then cropped and edited them like how human professional photographers would.
And again, the results speak for themselves.
Google approached the problem by relying to only a collection of professional quality photos, with before/after image pairs or any additional labels. The computers then break down the aesthetics into multiple aspects automatically, with each learning with negative samples generated by a couple image operation. The AI then enhanced the photo on its composition, saturation/HDR with lightning that is fast and separable.
The photos are cropped with saturation and HDR strength enhanced. Then a mask is applied to them for a concept of dramatic lightning. A traditional image filter was used to generate the negative training examples for saturation, HDR and composition. The negative examples were them generated by applying a combination of image filters that modify brightness randomly on professional photos, degrading their appearance.
For the training, the researchers used a generative adversarial network (GAN), where a generative model creates a mask to fix lighting for negative examples.
Google then showed a bunch of these photos, along with others taken from various sources, and asked several professional photographers for opinions. According to them, about 40 percent of the submissions were perceived as being created by "semi-pro" or "pro"-level photographers.
What this means, AI is somehow capable to applying contextual meaning and adjustments in different parts of each photographs, making them to have dramatic lightning and more compelling images - as opposed to just applying filters or adding a predicable vignette.
This work was done by Hui Fang and Meng Zhang from Machine Perception at Google Research.