Lightweight Detection and Counting of Maize Tassels in UAV RGB Images

By integrating unmanned aerial vehicle (UAV) remote sensing with advanced deep object detection techniques, it can achieve large-scale and high-throughput detection and counting of maize tassels. However, challenges arise from high sunlight, which can obscure features in reflective areas, and low sunlight, which hinders feature identification. Existing methods struggle to balance real-time performance and accuracy. In response to these challenges, we propose DLMNet, a lightweight network based on the YOLOv8 framework. DLMNet features: (1) an efficient channel and spatial attention mechanism (ECSA) that suppresses high sunlight reflection noise and enhances details under low sunlight conditions, and (2) a dynamic feature fusion module (DFFM) that improves tassel recognition through dynamic fusion of shallow and deep features. In addition, we built a maize tassel detection and counting dataset (MTDC-VS) with various sunlight conditions (low, normal, and high sunlight), containing 22,997 real maize tassel targets. Experimental results show that on the MTDC-VS dataset, DLMNet achieves a detection accuracy AP50 of 88.4%, which is 1.6% higher than the baseline YOLOv8 model, with a 31.3% reduction in the number of parameters. The counting metric R2 for DLMNet is 93.66%, which is 0.9% higher than YOLOv8. On the publicly available maize tassel detection and counting dataset (MTDC), DLMNet achieves an AP50 of 83.3%, which is 0.7% higher than YOLOv8, further demonstrating DLMNet's excellent generalization ability. This study enhances the model's adaptability to sunlight, enabling high performance under suboptimal conditions and offering insights for real-time intelligent agriculture monitoring with UAV technology.