Benchmarking AI architectures for circularity in textiles: fiber recognition with CNNs, ViTs, and hybrid models.

The explosion of textile waste requires high-efficiency, automated sorting systems to support large-scale recycling. Near-infrared (NIR) spectroscopy combined with Convolutional Neural Networks (CNNs) is a proven technique for non-destructive fiber identification. However, the effectiveness of CNNs in processing spectral data is often constrained by their architectural focus on local feature extraction. Advancements in Artificial Intelligence have generated novel architectures, such as Vision Transformers (ViT), which weigh the importance of all input elements relative to one another, effectively capturing long-range dependencies within a sequence, providing a promising approach for analyzing NIR spectral data derived from post-consumer textiles. This study aims to compare CNN (AlexNet, EfficientNetV2), ViT (ViT-base, ViT-large, DeiT, and BEiT), and Hybrid (ConvNeXt and Swin Transformer) architectures for classifying textile fibers, including cotton, nylon, and polyester. The results demonstrate ConvNeXt's superior performance, achieving the highest overall accuracy for textile fiber classification and indicating its potential to improve automated textile recycling systems.