Abstract: To address the issues of dense wheat ear distribution, variable posture angles in farmland scenarios, and the high parameter count and computational complexity of some existing detection models that made them difficult to deploy on mobile devices, this study proposed a lightweight improved model, YOLO-BRFA, for rotated wheat ear object detection, with YOLO11n-OBB as the baseline. In this method, BiFPN was introduced into the neck network to achieve bidirectional multi-scale feature fusion and reduce redundant computation. The C3k2_RFCBAMConv module was constructed in the backbone network to enhance the extraction of fine-grained features and suppress background interference. Meanwhile, the conventional convolutional downsampling layer was replaced by the lightweight ADown downsampling module, which reduced complexity while preserving fine-grained textures and inclined edge information as much as possible. Compared with the baseline model, the improved model increased precision, recall, and mAP₅₀ by 1.1, 2.1, and 1.4 percentage points, respectively, while reducing the parameter count, model file size, and computational cost by 38.1%, 33.3%, and 13.6%. The proposed model improved wheat ear detection performance while maintaining a lightweight design, and was suitable for deployment on resource-constrained edge computing devices.

Key words: YOLO11n-OBB, rotated object detection, BiFPN, C3k2_RFCBAMConv, ADown