FINE: Fabric Identification and loom-origin Evaluation, a Benchmark Dataset for Silk Classification and Loom Type Identitification

In this paper, we propose a benchmark dataset for categorization of fabric into two tasks: i) Classification of a given material into Silk versus Non-silk. ii) Identification of viewing method-Handloom or Powerloom. In the textile domain, due to high mislabelling and imitation, identifying Silk versus Non-silk fabrics and understanding the weave methodology is important for ensuring quality, cultural preservation, and catering to consumer preference. Since these are linked with economic value, material quality and consumer trust, misclassification can impact artisan income, the certification process and policy implementation; thus, protecting fair pricing and trust among buyers is necessary for buyers and industry stakeholders. The existing vision-based dataset are largely focused on fashion garments, defect detection, pattern garments, or addressing single-class studies. Hence, there is an absence of an RGB dataset that captures fabric weave geometry and contains samples from broader categories that jointly addresses two major tasks of textiles. To address these gaps, we introduce Fabric Identification and Loom Origin Evaluation FINE, a dataset captured using DSLR. The fabrics used are colourless and without prints to ensure that the dataset consists of images which are focused on the warp-weft structure of the fabric, and while training, the models are not affected by colour, patterns, prints or motifs during evaluation. We benchmark FINE using five state-of-the-art models: ResNet-50, EfficientNet-B0, ConvNeXt-Tiny, ViT-B/16, and Swin-T—and observe that transformer architectures outperform convolutional baselines, with ViT-B/16 achieving 98.37% accuracy for silk classification and 73.93% for loom origin. These results establish FINE as a strong benchmark for advancing textile informatics and industrial fabric authentication.