人工智能
计算机科学
标杆管理
特征(语言学)
模式识别(心理学)
机器学习
卷积神经网络
深度学习
人工神经网络
钥匙(锁)
特征提取
计算机视觉
训练集
出处
期刊:
日期:2026-03-27
卷期号:: 1-6
标识
DOI:10.1109/cacml68972.2026.11506628
摘要
Accurate classification of biodegradable fabrics is critical for sustainable textile recycling and quality control. Traditional supervised CNNs struggle with domain shifts from varying camera distances and image acquisition conditions, leading to significant performance drops. While self-supervised Vision Transformers promise domain-invariant representations, their edge over supervised CNNs in cross-domain fabric classification is unclear. This study compares models on the Biodegradable Fabrics Dataset of six classes: Abaca, Cotton, Hessian, Linen, Silk and Wool captured at “Near”, “Mid” and “Far” imaging configuration. Supervised models of (ResNet-50, EfficientNet-B3, ViT-Small/16) and self-supervised DINOv2 ViTSmall/14 were evaluated in a source-only domain generalization setting. All use linear probing on “Near” images, testing on all domains without fine-tuning. DINOv2 was also assessed via KMeans clustering. On “Far” domain, EfficientNet-B3 achieves 93.67 % accuracy, outperforming others. Self-supervised ViT beats supervised ViT by 5.67 percentage points, as discriminative self-distillation yields more domain-invariant features than ImageNet pre-training. DINOv2 achieves the best clustering: ARI 0.932, NMI 0.943, showing superior semantic separability across all models. Learning paradigm enhances domain robustness within the ViT family, while architecture and paradigm interact to determine absolute performance. These findings suggest that while CNNs remain superior for fixed-condition accuracy, selfsupervised Transformers offer a label-efficient pathway for discovering fabric categories unconstrained environments.
科研通智能强力驱动
Strongly Powered by AbleSci AI