Microscopic classification of cashmere, wool, and modified fibers using attention-guided feature fusion

Accurate identification of cashmere, wool and their modified variants is important for textile quality inspection, product authentication and anti-counterfeiting. However, their similar microscopic morphologies make fine-grained classification difficult, particularly when discriminative information is carried by subtle scale-edge patterns and high-frequency surface textures. To address this problem, this study develops a frequency-aware attention-guided cross-layer feature fusion network based on EfficientNet-B0. The squeeze-and-excitation units in the MBConv blocks are replaced with a Global Attention Mechanism (GAM) to preserve joint channel-spatial interactions, while a Laplacian-guided high-pass branch enhances shallow texture responses before adaptive fusion with deep semantic features. Under the fixed image-level test protocol, three independent training runs achieved an accuracy of 98.33 ± 0.42%, a macro-recall of 98.33% and a macro-F1 of 98.33%. To assess generalization to unseen physical specimens, five-fold specimen-grouped cross-validation was additionally performed. Each fold held out 24 specimens, and every one of the 120 specimens served as an unseen test specimen exactly once. The resulting image-level accuracy was 97.56 ± 0.52%. For specimen-level prediction, probabilities from all SEM images of the same specimen were averaged; all 120 out-of-fold specimen-level predictions were correct under the current acquisition conditions. Frequency-domain and feature-energy analyses further support the role of high-frequency information in this task. Controlled perturbation tests show that the model remains stable under moderate brightness, contrast, mild blur and small rotations, while stronger blur, noise and resolution degradation cause larger performance losses. These results support the proposed design for fine-grained microscopic fiber classification within the current acquisition domain.