Optimized Yolov5s-Im for real-time apple flower detection in drone-based pollination

Abstract

• Developed a lightweight optimized YOLOv5s-Im model for precise apple flower detection. • Achieved robust real-time performance on a drone-based pollination system, with more pollination attempts. • Validated YOLOv5s-Im through practical deployment across diverse resource-constrained platforms. • YOLOv5s-Im performance is superior to mainstream models. As traditional pollinators face increasing threats from climate change, the development of robotic pollination technology has become imperative, with apple flower detection emerging as a critical component of the technology. Deep learning (DL) advancements present novel methods in enhancing apple flower detection efficiency. However, deploying in real time on resource-constrained drone platforms demands a balance between computational efficiency and accuracy. To address this challenge, this study introduces an improved you-only-look-once version 5 small (YOLOv5s-Im) model by improving the original YOLOv5s architecture, using MobileNet version 3 as the backbone and GhostNet as the neck. This study then validated the YOLOv5s-Im performance by deploying it in real time on a drone platform designed for apple flower pollination. YOLOv5s-Im achieved an 88% detection accuracy and averaged 41.6 pollination attempts per 3-minute flight across five tests, significantly outperforming YOLOv5s and YOLOv5s with Transformers (YOLOv5s-T) as backbone (fewer than 10 attempts), due to its 2 FPS inference speed versus their 0.05 FPS. Control tests with lightweight models YOLOv5s with ShuffleNet version 2 (YOLOv5-Sh-V2) and YOLOv5s with MobileNet version 2 (YOLOv5s-Mb-V2) as backbones, averaged 37.8 and 30.6 attempts per flight, respectively, with accuracies of 80% and 82% mAP and detection speeds of 1.0 FPS and 0.7 FPS, further confirming YOLOv5s-Im’s superior balance of accuracy and efficiency. Its robust accuracy (84%-88%) across diverse conditions—clear light (88%), afternoon settings (86%), angled views (87%), and low-light shadows (84%)—demonstrates reliability in varied orchard environments. Compared to YOLOv5s, YOLOv5s-T, YOLOv7, YOLOv8, and Faster-R-CNN, YOLOv5s-Im excels with precision (90.6%), recall (87.7%), mAP50 (91.2%), and F1-score (89.42%), while reducing GFLOPS by 89% and model size by 85%, achieving high frame rates (227 FPS on NVIDIA RTX 4060 Ti, 22 FPS on Jetson Xavier, 4.56 FPS on Intel NUC11TNKi3). These results make YOLOv5s-Im an effective solution for real-time apple flower detection under natural lighting conditions in drone-based pollination systems.