Application of Data Augmentation for Accurate Attractiveness Estimation for Food Photography

This research aims to develop a data augmentation framework in order to improve the attractiveness estimation accuracy for food photography. Machine learning-based methods require numerous food images accompanied with their attractiveness for learning the relationship between image features and the attractiveness. To efficiently obtain such food images, we apply data augmentation; the proposed method applies four kinds of image transformations: rotation, scaling, shifting, and random noise addition to the original images accompanied with their attractiveness. The key idea here is to apply the image transformations within a parameter space in which the attractiveness of the transformed image can be regarded as the same as that of the original one. By this way, we can obtain a large number of images accompanied with their attractiveness without any additional subjective experiments. Experimental results showed the effectiveness of the proposed method framework.