A unified EfficientSwinB-based framework for medicinal plant recognition and safflower germplasm classification

The accurate recognition of medicinal plants and fine-grained crop genetic resources is essential for pharmacology, precision agriculture, and biodiversity conservation. However, leaf-based plant recognition remains challenging due to complex backgrounds, illumination variations, and subtle inter-class differences. To address these challenges, this study investigates plant image recognition at two levels of granularity. For general medicinal plant recognition under unconstrained outdoor conditions, we propose EfficientSwinB-SE, an enhanced Swin Transformer-based model that incorporates structured pruning and a squeeze-and-excitation module to improve efficiency and discriminative capability. For fine-grained classification of safflower germplasm varieties under controlled laboratory conditions, we further develop OLSF-EfficientSwinB-SE, a task-specific extension optimized for distinguishing closely related varieties. Experimental results demonstrate that EfficientSwinB-SE consistently outperforms SwinB and other state-of-the-art models, such as VGG11_bn, MobileNetV2, DenseNet, EfficientNetB4, ViT, ConvNeXt, and DeiT, in medicinal plant recognition, achieving accuracies of 99.75% on the Indian Medicinal Plant dataset and 97.70% on the Indonesian Medicinal Plant dataset. Meanwhile, OLSF-EfficientSwinB-SE achieves superior performance in fine-grained safflower germplasm classification with an accuracy of 91.63%. These findings confirm the effectiveness of the proposed framework for multi-level plant recognition, with practical applications in medicinal plant identification, germplasm management, and intelligent agricultural systems.