DFt: A Diffusion Transformer with Dual-Frequency Scaling Tanh for Few-Shot Wool and Cashmere Fibers Detection

The accurate identification of wool and cashmere fibers is crucial for textile quality, yet hindered by their visual similarity and the scarcity of annotated microscopic images. While deep learning techniques offer a high-accuracy, cost-effective identification technique, they are still limited by challenges arising from insufficient data scale and the high acquisition costs of obtaining high-quality fiber images. Although data augmentation partially alleviates data scarcity challenges, classical image augmentation techniques remain limited in generating high-fidelity, diverse fiber images, particularly in low-texture regions. This paper proposes a Transformer-based frequency-domain adaptive diffusion model incorporating a novel Dual-Frequency Scaling Tanh layer. The proposed method was evaluated on 1,845 high-quality fiber images (15 classes) provided by a National Fiber Quality Supervision and Inspection Center. The proposed approach demonstrates superior performance, outperforming all state-of-the-art baselines. Compared to classical image augmentation methods, our approach achieves performance gains of +0.02 F1-Confidence and +1.88% [email protected] in detection tasks, along with +0.05 F1-score, +4.83% precision, and +2.51% recall in classification tasks. These improvements demonstrate the method’s efficacy in generating high-fidelity, diverse data samples for intelligent fiber detection, while establishing an effective optimization strategy for deep learning model training. The source code is publicly available at https://github.com/zld-make.