Detection and Classification of Plant Fibers Considering Inter-Class and Intra-Class Feature Variability

Plant fibers, as a critical natural material, are extensively used in aerospace, composites, and bio-based engineering. Accurate characterization of fiber types and microstructures is essential for assessing raw material quality and controlling product performance. Conventional identification methods rely on laboratory microscopy and manual expert analysis, which are time-consuming, subjective, and unsuitable for automated, high-precision measurement systems. Microscopic images of plant fibers present challenges, including complex morphology, high inter-class similarity, significant intra-class variability, and limited sample availability, complicating automated feature-based classification. Existing convolutional neural network (CNN) approaches often fail to balance global structural and local textural features, while Transformer-based methods inadequately capture fine-grained local patterns. Hybrid models that attempt to combine these features frequently lack effective feature-fusion mechanisms, limiting their ability to exploit complementary information. To address these limitations, this work proposes CAFR-Net, a plant fiber image classification framework that integrates global and local features via dedicated structural and texture feature-extraction modules, combined with a cross-attention feature-interaction mechanism. The model enables collaborative feature modeling across multiple representation spaces, enhancing classification of complex fiber morphology. Evaluation on a dataset of 914 microscopic images across 32 fiber classes, using 5-fold cross-validation and an independent test set, demonstrates that CAFR-Net achieves a classification accuracy of 89.30% with a training-validation gap of 1.35% points, indicating strong generalization performance for small-sample measurement datasets. These results highlight the potential of CAFR-Net for automated, high-precision plant fiber characterization and quality assessment.