Test result of the algorithm. (Screenshot via Tencent Youtu Lab)
What will you do after taking a selfie? Directly post it online or beautify it using photo-editing apps first? As virtual makeup apps gain popularity among both males and females, Chinese developers are working on a way to remove makeups from over-edited photos.
During the International Conference of Computer Vision (ICCV) held from October 22 to 29 in Venice, Italy, Tencent Youtu Lab, an artificial intelligence (AI) research lab under the Chinese tech giant Tencent, released a new algorithm through a paper submitted to ICCV that can reverse over-edited deceiving pictures.
Dubbed "makeup-go," the algorithm is claimed to be the first of its kind in the world, and it focuses on tackling the two issues of skin smoothing and skin color change.
One of the difficulties mentioned by the lab is that every virtual makeup software performs quite differently, which make it impossible to learn all operations that can smooth skin, suppress wrinkle and freckle, and adjust tone.
Therefore, the team designed a new network structure to reverse pictures "blindly." Learned from data of different software, the network can restore beautified images without knowing the exact algorithms.
Also, instead of directly regressing the pictures towards the unedited version, the network regresses levels of principal components of the edited parts separately. As a result, subtle detail information won't be ignored.
Makeup-go doesn't handle geometric face-lifting at the moment, the team said in the paper.
