Performance Evaluation of YOLOv8 to YOLOv11 for Accurate Detection and Classification of Stomata in Microscopic Images of Herbal Plants

Abstract

This study benchmarks the latest generation of YOLO (You Only Look Once) deep learning architectures for microscopic image analysis of plant stomata. This study implemented and compared four YOLO variants (YOLOv8, YOLOv9, YOLOv10, and YOLOv11) to create an automated computer vision system capable of detecting and classifying four distinct stomatal morphology types in microscopic images of herbal plants. The dataset consisted of 480 images from four different herbal plants. Performance evaluation using mean Average Precision (mAP@50 and mAP@50-95), precision, recall, and validation loss metrics revealed that YOLOv9 and YOLOv11 consistently outperformed other variants, achieving mAP@50 values exceeding 0.95. YOLOv11 demonstrated superior precision (0.93348), while YOLOv9 exhibited the best recall (0.89876) and lowest validation loss. Notably, YOLOv11 demonstrated significant computational efficiency advantages, with the fastest processing time (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$4,542.97$</tex> seconds), making it particularly suitable for high-throughput applications. Interestingly, the findings reveal that newer YOLO versions do not always correspond to performance improvements, as YOLOv9 outperformed YOLOv10 in almost all evaluation metrics. This computational approach eliminates the need for manual expert analysis, providing a scalable and high-throughput solution for object detection and the identification of microscopic biological structures. The proposed system offers significant advancement in automated image analysis for botanical research, with potential applications in digital phenotyping, plant species identification, and quality assessment of herbal plants through computer vision techniques.

Access to Document

Other files and links

• Link to publication in Scopus

Fingerprint