A Lightweight Neural Network for Accurate Rice Panicle Detection and Counting in Field Conditions

Abstract

Monitoring rice spikelet yield is crucial for ensuring food security, but manual observations are tedious and subjective. Deep learning approaches for automated counting often require high device resources, limiting their applicability on low-cost edge devices. This paper presents the Rice Lightweight Feature Detection Network (RLFDNet). RLFDNet designed for the field of computer vision, features a lightweight encoder and decoder, effectively decoding shallow and deep information within its neural network architecture. Innovative designs including dense feature pyramid network, reinforcement learning guidance, attention mechanisms, dynamic receptive field adjustment, and shape feature fusion enable outstanding performance in object detection and counting, even with low-resolution images. Across different elevations, ranging from 7m to 20m, RLFDNet demonstrates significantly superior accuracy and inference efficiency compared to other advanced object detection methods. With a parameter count of only 4.40 million, it achieves an impressive frame rate of 80.43 FPS on a GTX1080Ti GPU, meeting real-time application requirements on inexpensive devices. RLFDNet's exceptional performance is further highlighted by an MAE of 1.86 and an R2 of 0.9461, along with an average precision of mAP@0.5 reaching 0.91. These results underscore RLFDNet's capability as a potent and reliable visual tool for agricultural practitioners, offering promising prospects for future research endeavors.