2D FFT and AI-Based Analysis of Wallpaper Patterns and Relations Between Kansei

The human ability to texture pattern recognition is very high and precise. Although details of texture significantly affect Kansei, grasping texture features in quantitative methods have been difficult. We have analyzed wallpaper texture patterns and Kansei, with Principal Component Analysis, 2-dimensional FFT (Fast Fourier Transfer) and Convolutional Neural Networks. Principal Component Analysis showed the Kansei structure on wallpapers. 2D FFT results are used for revealing specific relations between spectrum features and Kansei evaluation. Convolutional neural networks have learned to be Kansei visual recognition system and integrative feature analyzer. 2DFFT was used to analyze 3 significant samples that differ only on texture. Square staggered texture, small and large rhombus textures have different FFT patterns. The planar frequency patterns suggest different Kansei perceptions. CNN has learned as "transfer learning" based on pre-trained networks. Pattern perception and relations between Kansei structures were successfully learned. Interpolations for unlearned patterns were also investigated.