Optimized ResNet18 for few-shot fabric texture image recognition: enhancing accuracy in digital textile applications

Accurate fabric texture image recognition is essential to the digital transformation of the textile and apparel industry, enabling applications such as virtual garment design, AI-driven fashion development, and digital rights management. Most current methods rely on deep learning models like ResNet, which extract visual features through convolutional operations mimicking human visual cognition in a hierarchical manner. However, these approaches typically require large, well-labeled datasets for effective training. In practice, acquiring such datasets is challenging. Fabric textures vary significantly due to differences in materials, manufacturing techniques, and intended functions. This complexity, combined with limited data sharing, a lack of standardized classification systems, and industry confidentiality, restricts access to large-scale public datasets. As a result, few-shot learning has become a pressing need in accurate fabric texture image recognition. The standard ResNet18 model begins with a large 7 × 7 convolutional kernel, designed for coarse feature extraction. While effective in general image tasks, this structure often fails to capture fine-grained texture details in few samples scenarios and increases the risk of overfitting. To overcome these limitations, we propose an optimized ResNet18 architecture tailored for few samples fabric texture image recognition. Our approach replaces the initial 7 × 7 kernel with three sequential 3 × 3 kernels, promoting more detailed and layered feature extraction. We evaluate the proposed model on a dataset comprising eight typical women’s shirt fabric textures. Experimental results show that the optimized model achieves higher recognition accuracy and faster training convergence compared to the baseline ResNet18. This work contributes to more efficient and scalable digital solutions in textile and apparel technologies.