MedPlantNet: Spatial Attention and Residual Multi-Scale Feature Fusion Integrated Deep Network for Medicinal Plants Recognition

Recognizing medicinal plants that are used to cure or prevent ailments has gained substantial research interest owing to challenges in manual recognition. The increasing demand for automated recognition system regarding medicinal plants has led to an emphasis on employing leaf images that are readily accessible and provide extensive information. The requirement of efficient medicinal plants recognition system is further emphasized by the significant influence of medicinal plants in the healthcare sector. Incorporating ensemble operation, residual operation, and spatial attention, this paper presents MedPlantNet, a robust deep neural network architecture for automated medicinal plants recognition. Additionally, a residual multi-kernel convolution block guided by spatial attention in conjunction with pretrained features has been introduced. The "Bangladeshi medicinal plant" dataset and the "BDMediLeaves" dataset, which are the most recently available datasets on medicinal plants found in Bangladesh, were employed to evaluate the performance of the proposed automated medicinal plants recognition system architecture. The proposed architecture has recognized medicinal plants with an accuracy of 98.40% and 98.51% on the "Bangladeshi medicinal plant" and "BDMediLeaves" datasets, respectively. Our proposed architecture has outperformed previously proposed architectures with fewer epochs and lower computational time without any rigorous image preprocessing.