MSAF-Net: Multi-Scale Attention Fusion for Fine-Grained Medicinal Plant Recognition Under Near-Duplicate-Audited and Cross-Dataset Evaluation

Recognising medicinal plant species from leaf images is a fine-grained problem: silhouette and venation provide complementary evidence at different feature scales. We propose MSAF-Net, which taps three stages of an ImageNet-22k pre-trained ConvNeXt-Tiny backbone, refines each stage with convolutional block attention, lets the scales interact through spatial cross-scale attention over 4×4 token grids, and fuses them with an image-conditioned gate. The fixed split of MPBD-18 (18 Bangladeshi species, 17,383 field images) is audited for near-duplicates: 0.86% of its test images have a training near-duplicate, and removing them leaves the ranking unchanged. The second benchmark is split group-aware via perceptual hashing. On MPBD-18, MSAF-Net reached 99.94 ± 0.06% accuracy over three seeds, ranking first among 15 tuned models; it significantly outperformed four classical CNNs under Holm-corrected exact McNemar tests, while no significant difference was detected against the strongest modern backbones. Re-benchmarking every architecture from its ImageNet initialisation on an independent 80-species Indian dataset (6900 images) again placed MSAF-Net first (96.70% accuracy, Cohen’s κ 0.9666), significantly ahead of nine out of fourteen baselines; this measures architectural portability, not transfer of a trained model. A seed-level paired analysis supports the multi-scale fusion, whereas the finer design choices remain unresolved at this benchmark’s accuracy level.