MSFFAR-Net: A Multi-Scale Feature Fusion and Attention-Guided Residual Network for Bangladeshi Medicinal Plants Classification

The classification of plant species, particularly medicinal ones, has gained much research attention due to challenges in manual identification and the urgency to protect endangered species. The growing need for automated identification of medicinal plants has led to a focus on using leaf images, which provide abundant information and are easily accessible. The importance of medicinal plants in healthcare underscores the need for efficient identification methods, given the time-consuming reliance on experts. Vision-based systems offer fast and accurate recognition, benefiting researchers and the public in identifying these essential plant species. This work proposed a Multi-Scale Feature Fusion and Attention-Guided Residual Network (MSFFAR-Net) that incorporates pre-trained MobileNetv2 as feature extractor and a series of newly introduced Attention Residual Convolution Module to classify Bangladeshi Medicinal plants. Our model, validated through a 4-fold cross-validation, achieved high classification mean accuracy of 98.12% for BDMediLeaves and 96.18% for the Bangladeshi Medicinal Plant Dataset.