Visual Similarity Assessment for Product Aesthetic Properties Using Single Reference Training

Traditional product aesthetic evaluation requires extensive preference labeling for each design, limiting scalability and increasing costs. This study presents a novel single reference training approach that enables aesthetic preference assessment through visual similarity to a single highly preferred product. Using 30 home audio speakers, we collected aesthetic preference ratings from 44 participants on a 7-point Likert scale. Four deep learning approaches—Pre-trained CNN, Auto-encoder, Siamese Network, and Triplet Network—measured visual similarity between themost preferred reference products and remaining samples. Results demonstrated significant correlations between visual similarity and aesthetic preference: the Triplet Network achieved Pearson correlation coefficient <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">r</i> = 0.448 (<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">p</i> = 0.013), while Pre-trained CNN approach yielded <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">r</i> = 0.478 (<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">p</i> = 0.008). After filtering high-variance products, correlations substantially improved to <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">r</i> = 0.738 (<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">p</i> < 0.001). Principal component analysis of embedding vectors revealed interpretable aesthetic dimensions, with specific components significantly correlating with novelty (<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">r</i> = 0.51 − 0.61), harmony (<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">r</i> = 0.36 − 0.37), dynamics (<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">r</i> = 0.55 − 0.65), and complexity (<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">r</i> = 0.42 − 0.56). Interaction models of multiple principal components increased explanatory power (<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">R</i><sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> = 0.4065), demonstrating that aesthetic preferences emerge from complex relationships among visual features. The findings prove that single reference training effectively extracts interpretable aesthetic properties from embedding spaces without explicit labeling. Our results suggest that highly preferred products contain multiple aesthetic preference-related properties, and deep learning models can successfully extract features corresponding to these properties. This approach provides an alternative aesthetic assessment from labor-intensive individual evaluation to efficient similarity-based inference, offering a methodology with reduced data requirements for product design evaluation and enabling evaluation of new designs.