Recognition of terminal buds of densely-planted Chinese fir seedlings using improved YOLOv5 by integrating attention mechanism

• DOI: 10.3389/fpls.2022.991929
• Publication Date: 2022-10-10T06:52:10Z
• Authors: Zhangxi Ye; Qian Guo; Jiahao Wei; Jian Zhang; Houxi Zhang; Liming Bian; Shijie Guo; Xueyan Zheng; Shijiang Cao
• Keywords: Robustness

Accurate and timely information on the number of densely-planted Chinese fir seedlings is essential for their scientific cultivation intelligent management. However, in later stage cultivation, overlapping lateral branches among individuals too severe to identify entire individual UAV image. At same time, high-density planting nursery, terminal bud each seedling has a distinctive characteristic growing upward, which can be used as an identification feature. Still, due small size dense distribution buds, existing recognition algorithm will have significant error. Therefore, this study, we proposed model based improved network structure latest YOLOv5 identifying seedlings. Firstly, micro-scale prediction head was added original enhance model's ability perceive small-sized buds. Secondly, multi-attention mechanism module composed Convolutional Block Attention Module (CBAM) Efficient Channel (ECA) integrated into neck further focus key target objects complex backgrounds. Finally, methods including data augmentation, Test Time Augmentation (TTA) Weighted Boxes Fusion (WBF) were improve robustness generalization buds different growth states. The results showed that, compared with standard version YOLOv5, accuracy significantly increased, precision 95.55%, recall 95.84%, F1-Score 96.54%, mAP 94.63%. Under experimental conditions, other current mainstream algorithms (YOLOv3, Faster R-CNN, PP-YOLO), average also increased by 9.51-28.19 percentage points 15.92-32.94 points, respectively. Overall, attention accurately images provide technical support large-scale automated counting