Texture Synthesis Based on Aesthetic Texture Perception Using CNN Style and Content Features

Abstract

We propose a texture synthesis method that controls the desired visual impressions using CNN style features and content features. Diversifying user needs has led to the personalization of products according to individual needs. In the custom made garment service, users can select and combine fabrics, patterns, and shapes of garments prepared in advance to design garments that meet their tastes and preferences. Furthermore, controlling the visual impressions will enable the service to provide designs that better match the user’s preferences. In image synthesis, controllable texture synthesis have been performed with style and content, however, few previous studies have controlled images based on impressions (including aesthetics). In this study, we aim to synthesize textures with desired visual impressions by using style and content features. For this purpose, we first (1) quantify the affective texture by subjective evaluation experiments and (2) extract style features and content features using VGG-19 from pattern images for which evaluation scores are assigned. The explanatory variables are style and content features, and the objective variables are evaluation scores. We construct an impression estimation model using Lasso regression for each of them. Next, (3) based on impression estimation models, we control the visual impressions and synthesize textures. In (2), we constructed highly accurate visual impression estimation models using style and content features. In (3), we obtained synthesis results that match human intuition.