Effective Multi‐Layer Semantic Fusion Method for Tibetan Medicinal Plant Detection

Tibetan medicinal plants possess significant therapeutic value due to their unique pharmacological properties. However, manual identification remains challenging owing to their remote and often inaccessible growing environments. To overcome these detection difficulties, a multi‐layer semantic fusion (MLSF) method is proposed for automated Tibetan medicinal plant recognition. The proposed method integrates a cascaded fusion network within the feature extraction backbone to enhance multi‐scale feature representation, addressing the substantial scale variations inherent in dense plant scenes. Additionally, a dedicated small‐object detection layer is incorporated into the feature fusion network to improve detection accuracy for diminutive specimens. To further optimise feature integration, an asymptotic feature pyramid network is employed at lower semantic levels, enabling progressive multi‐level feature fusion. Finally, two penalty terms are introduced into the regression loss function to mitigate missed detections caused by plant occlusion. Experimental evaluations on our custom dataset demonstrate that MLSF achieves a mean average precision (mAP)@.5 improvement from 85.7% to 87.6% and an mAP@.5:.95 increase from 70.1% to 71.7%, outperforming the baseline YOLOv8 model. These results underscore the efficacy of our approach, highlighting its potential for broader applications in dense vegetation detection tasks.