Mobile Device Facial Beauty Prediction using Convolutional Neural Network as Makeup Reference

Abstract

Beauty or good looks is one aspect of human attraction. Highly attractive people have an 8 to 12 % higher success in negotiations. One of the things that can be used to increase attractiveness is makeup or cosmetics. Makeup can be used alone or through the services of a Makeup Artist (MUA). MUA services are often used during important and once-in-a-lifetime events such as weddings. The difference in beauty before and after using this makeup cannot be measured but only seen and assessed manually. One application of machine learning in this field is Facial Beauty Predictions (FBP). FBP measures the value of a person's facial beauty or good looks. This paper investigated and implemented FBP on mobile devices, which makes it easier to assess anytime and anywhere. We investigated FBP models using five CNN architectures, including EfficientNetB0, VGG-16, ShuffleNet, MobileNet, and Inception. After evaluation, we choose MobileNetV2 CNN architecture as the primary model for mobile application FBP. Experiments on the FBP dataset show that MobileNetV2 achieves Pearson Correlation 0.7893 with only 10.5 MB for model file size. The model implemented for mobile devices works well in 13 different tests.