A field rice panicle detection model based on improved YOLOv11x.

Rice serves as the staple food for over 50% of the world's population, making its yield prediction crucial for food security. The number of panicles per unit area is a core parameter for estimating rice yield. However, traditional manual counting methods suffer from low efficiency and significant subjective bias, while unmanned aerial vehicle (UAV) images used for panicle detection face challenges such as densely distributed panicles, large scale variations, and severe occlusion. To address the above challenges, this paper proposes a rice panicle detection model based on an improved You Only Look Once version 11x (YOLOv11x) architecture. The main improvements include: 1) Introducing a Bi-level Routing Attention (BRA) mechanism into the backbone network to improve the feature representation capability for small objects; 2) Adopting a Transformer-based detection head (TransHead) to capture long-term spatial dependencies; 3) Integrating a Selective Kernel (SK) Attention module to achieve dynamic multi-scale feature fusion; 4) Designing a multi-level feature fusion architecture to enhance multi-scale adaptability. Experimental results demonstrate that the improved model achieves an mAP@0.5 of 89.4% on our self-built dataset, representing a 3% improvement over the baseline YOLOv11x model. It also achieves a Precision of 87.3% and an F1-score of 84.1%, significantly outperforming mainstream algorithms such as YOLOv8 and Faster R-CNN. Additionally, panicle counting tests conducted on 300 rice panicle images show that the improved model achieves R