Enhanced YOLOv11 with Dual Attention Mechanisms for Fine-Grained Stomata Morphology Classification

Abstract

Accurate classification of stomatal morphological types is essential for botanical taxonomy and pharmacognostic authentication of herbal plants; however, no prior system has simultaneously detected and classified all four types – Anomocytic, Diacytic, Paracytic, and Graminoid – within a unified multi-instance detection pipeline. This study proposes an Enhanced YOLOv11 architecture integrating the Convolutional Block Attention Module (CBAM) at two successive C2f backbone junctions for progressive feature refinement and Adaptive Spatial Feature Fusion (ASFF) at each of the three detection scales to resolve inter-scale feature conflicts. The model was trained on a dataset of 6259 microscopic images and 74424 annotated instances across four morphological types and eight herbal plant species. Ablation experiments confirm that the dual-attention configuration achieves an mAP50 of 0.895 and an F1-score of 0.846, with a nonlinear synergy gain of 7.53%, whereby the CBAM-ASFF model achieves a lower inference latency (114.2 ms) than the ASFF-only variant (121.7 ms) despite the near-identical computational cost (9.091 vs. 9.088 GFLOPs). When benchmarked against CA-YOLOv5, StomaYOLO, and YOLOv8s retrained on the same dataset, the proposed model achieved the highest mAP50 and F1-score while maintaining the lowest parameter count (3.371 M) among all evaluated architectures.

Access to Document

Other files and links

• Link to publication in Scopus
• Open Access Version Available

Fingerprint