Facial Image Aesthetics Prediction with Visual and Deep CNN Features

Facial Image Aesthetics Prediction with Visual and Deep CNN Features
Mohamed Selim, Tewodros Amberbir Habtegebrial, Didier Stricker
Irish Machine Vision and Image Processing Conference Irish Machine Vision and Image Processing Conference (IMVIP-17), August 30 - September 1, Maynooth, Ireland

Abstract:
Large number of images that has persons are being uploaded to the Internet, at a very high rate. However, they vary in quality and aesthetics. These variations affect the performance of the facial images analysis algorithms. This fact poses an interesting question: Can we predict the aesthetics of the facial image in stills?. In this work, we introduce a framework that uses deep face representations from CNNs and other visual features to tackle the problem. We evaluated our algorithms on large scale datasets of persons. Regarding the aesthetics, we used collected portraits from the AVA dataset, as well as the Selfie dataset. We thoroughly evaluated our algorithm. Moreover, we outperformed the state-of-the-art in aesthetic prediction in portrait images as we achieved accuracy of 84% while the state-of-the-art achieved 64.25% by using deep representations from our AestheticsNet combined with visual features
Keywords:
Aesthetics, CNN, Deep Learning, Facial, Selfie, Portraits