Morphology-prior discriminative fusion network for SEM-based classification of cashmere, wool, and modified wool fibers

Scale-removal and surface-smoothing treatments weaken the distinctive cuticle structures of wool fibers, including scale edges, surface roughness, and boundary fluctuations. As a result, modified wool exhibits an intermediate scanning electron microscopy (SEM) morphology between natural wool and cashmere, which obscures decision boundaries and reduces interclass separability in image-based classification. To address this ambiguity, we propose an improved MobileNetV3-Large framework that combines morphology-guided representation and discriminative feature refinement. Explicit morphological descriptors are extracted to preserve weakened cuticle-related cues, while MobileNetV3-Large learns deep texture representations from SEM images. To make the deep representation more consistent with subtle morphological differences, a discrimination aware mechanism (DAM) first filters ambiguous convolutional responses using class-proxy guidance, and cross-modal attention (CMA) then uses the retained discriminative features to guide the fusion of complementary morphological descriptors. This design enhances category separation among cashmere, wool, and modified wool. Experiments on a self-constructed SEM image dataset yielded 98.52% accuracy for the fusion model, surpassing baseline MobileNetV3-Large by 5.56%. These results affirm the fusion model's superior performance and the feature integration strategy's role in bolstering discriminability for fiber identification.