Research on detection of wheat tillers in natural environment based on YOLOv8-MRF

Abstract

To bolster agricultural efficiency and precision, this study introduces the YOLOv8-MRF model (multi-path coordinate attention, receptive field attention convolution, and Focaler-CIoU-optimized YOLOv8), a groundbreaking advancement in automated detection of wheat tillers. This model transcends traditional manual methods prone to subjectivity and inefficiency. This approach integrates an enhanced multi-path coordinate attention (MPCA) mechanism within the backbone network, capturing multi-scale features and significantly elevating tillers recognition. The innovative replacement of the CSPDarknet53 to 2-Stage FPN (C2F) module with receptive field attention convolution (RFCAConv) addresses parameter-sharing limitations, accentuating feature significance, and amplifying network performance. Coupled with the Focaler-CIoU loss for superior detection accuracy, YOLOv8-MRF outperforms RTDETR, YOLOv5, YOLOv7, and YOLOv8 by impressive margins in mAP50, while operating with merely 11 % of the parameters of YOLOv7, achieving a detection precision of 91.7 %, and with enhancements of 2.5 % in precision, 5.5 % in recall, and 4.1 % in mAP50 over the original model. The experimental results demonstrate that this method can realize tillering detection under complex backgrounds, contributing to advancing intelligent farming practices for wheat. Importantly, the YOLOv8-MRF model not only achieves significant technological advancements but also shows strong potential in practical applications, providing an effective tool for agricultural automation and intelligence, which could become pivotal in the development of future precision agriculture technologies.