PIKAR: A Pixel-Level Image Kansei Analysis and Recognition System Based on Deep Learning for User-Centered Product Design

摘要

Can machines learn to perceive products like humans? Kansei Engineering has been developed to connect product design and human perception. While conventional Kansei Engineering Systems exhibit a high dependency on manual extraction of design elements and thereby are restricted to validity issues, we present a Pixel-level Image Kansei Analysis and Recognition (PIKAR) system that applies deep learning to extract and analyze the formation of human perception towards product designs automatically. Method validation is performed based on the evaluation of cosmetic packaging's kawaii. Two neural nets trained on 1,414 images, labeled by eight participants based on their perception of kawaii (1-5 Lik ert Scale), have achieved a better prediction than test persons. The implemented neuron analysis methodology for Kansei analysis points towards consistency with previous experimental kawaii studies and gives insight to individual differences. This work addresses the possibility of applying deep learning to support product design and user experience researches.