Lightweight and Robust YOLOv5s with MobileNetV3 and GhostNet for Precision Apple Flower Detection for Pollination Drones

The improved YOLOv5s model represents a meticulous endeavor to balance efficiency and accuracy in apple flower detection within natural environments for pollination drones. Through strategic modifications, leveraging a MobileNetV3 backbone and GhostNet head, the model achieves a significant reduction in computational complexity while maintaining commendable performance metrics. Comparative analysis positions it as a standout choice, bridging the accuracy gap between modified YOLOv5s with MobileNetV3 backbone and GhostNet head and YOLOv5s, YOLOv7 and YOLOv8s while outperforming alternatives like Faster R-CNN. Precision, recall, and mean Average Precision (mAP50) metrics underscore its efficacy, with precision and recall values reaching 90.6% and 87.7%, respectively, alongside a remarkable mAP50 score of 91.2%. Crucially, the reduction in GFLOPs and model size by approximately 89% and 85%, respectively as compared to YOLOv5s benchmark model, enhancing its deploy ability on constrained devices. Execution tests on embedded devices further validate its practicality, demonstrating impressive 227 FPS on NVIDIA GeForce RTX 4060 Ti GPU device and 20 Frames Per Second (FPS) ratings, particularly on resource-constrained platforms like the Jetson devices. Looking ahead, opportunities for fine-tuning under specific environmental conditions and extending applicability to diverse flora remain, alongside the exploration of edge computing optimizations to further streamline deployment. This modified YOLOv5s emerges as a robust solution, poised to address the demand for lightweight yet accurate computer vision models in real-world scenarios for pollination dronesS.