Rice spikelet flowering-state detection based on an improved YOLOv11n model.

Accurate identification of rice spikelet flowering status is critical for germplasm screening in hybrid rice programmes. However, automated detection remains challenging because rice spikelet targets are small, densely distributed, and subject to significant background interference. To address these challenges, this study proposes YOLO11n-ACDW, an improved YOLO11n-based model for rice spikelet flowering-status detection. Specifically, the model incorporates the ADown downsampling module to preserve high-frequency features of small targets, introduces a Convolution-Attention Fusion Module (CAFM) to enhance discrimination between targets and background, adopts the Detect_Efficient detection head to adapt to the scale characteristics of rice spikelet, and integrates the Wise-IoU v3 loss function to improve bounding-box regression accuracy. Experimental results showed that the improved model achieved a precision of 88.8%, a recall of 84.7%, and an mAP@0.5 of 84.0% on the test set, representing improvements of 8.0%, 3.4%, and 1.3% over the baseline YOLO11n, respectively. Compared with the mainstream detector RT-DETR, the proposed model improved these metrics by 9.2%, 6.1%, and 2.5%. In terms of model complexity, it also outperformed the other algorithms, with only 2.14 M parameters and a computational cost of 4.4 GFLOPs, while maintaining lightweight characteristics. These results provide technical support for flowering-window monitoring in hybrid rice breeding.