A novel mobile-captured fabric dataset for benchmarking cnn architectures in multiclass classification

Abstract

This report investigates automated fabric-type identification under realistic textile production conditions using smartphone images captured without controlled lighting, backgrounds, or camera positioning. A dataset of 3,354 images covering six fabric classes—canvas, twill, denim, RFD knits, fleece, and corduroy—was developed, with areal-density measurements ranging from 186 to 371 g/m². Six ImageNet-pretrained CNN models were evaluated using a common two-stage transfer-learning approach, with and without data augmentation. EfficientNetV2-B0 achieved the best performance, reaching 99.8% accuracy and 99.9% precision, recall, and F1-score with augmentation, while MobileNetV2 provided the fastest inference at 6.2 ms per image. Grad-CAM analysis showed that the models focused primarily on fabric surfaces rather than backgrounds or measurement labels, reducing concerns about shortcut learning. Overall, the study demonstrates that lightweight, compound-scaled deep learning models can accurately classify fabric types from ordinary smartphone images under uncontrolled, real-world textile production conditions.