Fusing Global and Local Information Network for Tassel Detection in UAV Imagery

Abstract

Unmanned aerial vehicles (UAVs), equipped with sensors, have made a significant impact in the field of agricultural analysis. Maize, being one of the most vital crops worldwide, is intricately linked to its yield and the growth of tassels. Leveraging UAV imagery for the automatic monitoring of maize tassels holds the potential to drive the development of intelligent maize cultivation. Current research methods, nevertheless, are limited and lack robustness. To address the challenge of tassel detection in UAV images, we propose an innovative network, termed FGLNet. This network models the backbone with a 16x down-sampling to retain richer pixel information and enhances performance by effectively fusing global and local information through weighted mechanisms. Moreover, the scarcity of tassel data presents a substantial constraint. In this article, we publicly release a new dataset, named the maize tassels detection and counting UAV (MTDC-UAV), featuring annotated bounding boxes, to advance research in the agricultural domain. Although tassel detection and counting in aerial images pose formidable challenges, our approach demonstrates remarkable accuracy in evaluations based on the MTDC-UAV dataset. It achieves a detection AP$_{50}$ of 0.837 and a counting R$_{2}$ of 0.9409, all while maintaining a parameter count of just 0.77 M. This level of performance considerably outperforms other state-of-the-art computer vision methods. Overall, this research not only introduces innovative concepts but also provides worthwhile references and a solid data foundation for future studies.

Keywords

• Computer vision
• detection and counting
• information fusion
• maize tassel
• unmanned aerial vehicle (UAV)