Skin beautification detection using sparse coding

In the past years, skin beautifying softwares have been widely used in portable devices for social activities, which have the functionalities of turning one's skin into flawless complexion. With a huge number of photos uploaded to social media, it is useful for users to distinguish whether a photo is beautified or not. To address this problem, in this paper, we propose a skin beautification detection method by mining and distinguishing the intrinsic features of original photos and the corresponding beautified photos. To this aim, we propose to use sparse coding to learn two sets of basis functions using densely sampled patches from the original photos and the beautified photos, respectively. To detect whether a test photo is beautified, we represent the sampled patches from the photo using the learned basis functions and then see which set of basis functions produces more sparse coefficients. To our knowledge, our effort is the first one to detect skin beautification. To validate the effectiveness of the proposed method, we collected about 1000 photos including both the original photos and the photos beautified by a software. Our experimental results indicate the proposed method achieved a desired detection accuracy of over 80%.