A novel product shape innovation design method based on Kansei Engineering and GAN model with limited sample data

Abstract

There is a growing emphasis on fulfilling emotional as well as functional needs in product design. However, the current product design process often relies solely on designers’ experience and knowledge, resulting in inconsistent design solutions. Using GANs for product design is a promising direction. In this paper, we propose a new method for product modelling based on Kansei Engineering and GAN under finite samples. First, we collected and processed product images to build a dataset for perceptual factor scoring and training the generative model. Next, we collected a set of Kansei words and analysed them using the SD method and TOPSIS to extract perceptual factors and finalise the Kansei words. Third, through expert grading, we established the mapping between Kansei words and sample images. We then fed the sample images, Kansei words, and associated scores into the model for training. After training, the model generates new concept patterns. Finally, we applied the entropy weight method to evaluate both the conceptual design scheme and the sample images. The results show that the generated images are more consistent with the perceptual vocabulary than the original sample images. Users can then leverage these images for product styling, thereby aiding their design process. Abbreviations: DCGAN: Deep Convolutional Generative Adversarial Networks; GANs: Generative Adversarial Networks; KE: Kansei Engineering; CNN: Convolutional Neural Network; SD: Semantic Differential; TOPSIS: Technique for Order Preference by Similarity to Ideal Solution; KJ: Kawakita Jiro Method; KENPI: Kansei Engineering with a Deep Learning; GRA: Grey Relational Analysis; PFBD: Probability Feature Based Discriminant; MAGAN: Multi-Agent Generative Adversarial Network; LSGAN: Least Squares Generative Adversarial Net- work; PD-GAN: Perceptual Discriminative Generative Adver- sarial Network; WGAN: Wasserstein Generative Adversarial Network; RSGAN: Relativistic Standard Generative Adversarial Network; EEG: Electroencephalography; KMO: Kaiser-Meyer-Olkin (Measure of Sampling Adequacy); EWM: Entropy Weight Method; MCDM: Multiple Criteria Decision Making; FID: Fréchet Inception Distance; URL: Uniform Resource Locator; SSIM: Structural Similarity Index. © 2025 Informa UK Limited, trading as Taylor & Francis Group.