Deep learning-based visual-tactile perception and quantitative rating of fabrics

This is the authors' abstract. We don't add key points for this paper.

In e-commerce clothing shopping, due to the absence of tactile feedback for consumers, the visual texture presented in product images often deviates from the actual tactile experience, leading to persistently high return rates. To explore the relationship between fabric image visual stimuli and human subjective perception, this study uses a self-developed dynamic image acquisition system for an exploratory, feasibility-driven investigation. By capturing dynamic image samples of representative fabrics in a controlled wind environment and combining them with subjective sensory evaluation data, the study builds a predictive mapping model based on the VGG16 deep learning architecture. This model establishes a strong connection between dynamic visual features and subjective perception ratings. The framework is rigorously trained and validated under 3, 5, and 7-level rating scales. The results show that the proposed model effectively predicts five tactile attributes, including softness and drapability. With the current sample size, the 3-level rating strategy performs best, with F1 scores for all attributes exceeding 0.65. This study confirms the technical feasibility of combining dynamic visual information with deep learning for fabric tactile evaluation. It provides an empirical foundation for the digital transformation of sensory data in the textile industry and offers valuable insights for improving consumer experience on e-commerce platforms and automating fabric quality ratings.