Abstract

Rapid and accurate counting of wheat ears is a key link in field yield estimation and breeding evaluation. The Global Wheat Head Detection (GWHD) dataset is widely used for the evaluation of deep learning models. However, as it covers complex field scenarios involving multiple environments and varieties, it poses a severe challenge to the generalization ability of detection models. To address the limitations of existing methods on the GWHD dataset, including poor adaptability to multi-scale targets, weak ability to capture features of tilted wheat heads, and inaccurate localization in dense occlusion scenarios, this paper proposes a lightweight improved model based on YOLOv11, namely Dynamic Fusion YOLOv11 (DF-YOLOv11). Firstly, a Dynamic Multi-Scale Fusion (DMS-Fusion) module is designed in the neck network, which effectively coordinates the extraction of detailed and semantic features of wheat heads at different scales through scale-adaptive grouped convolution and a dual-branch enhancement structure. Secondly, a Lightweight Direction-Aware Attention Module (Light-OrientedECA) is embedded in the backbone network. Utilizing directional pooling tailored to the angular characteristics of the GWHD dataset and depthwise separable convolution, it significantly improves the recognition ability of tilted wheat heads with only a slight increase in computational cost. Finally, based on the statistical characteristics of the GWHD dataset, dedicated anchor boxes are generated via K-means clustering, and the CIoU loss function is improved by introducing an occlusion-aware weight (Occlusion-CIoU), enhancing the model's robustness in bounding box regression under overlapping and occluded scenarios. Comprehensive experiments on the GWHD test set show that DF-YOLOv11 achieves an [email protected] of 87.3%, with only 1.57M parameters and 5.9 GFLOPs of computational complexity. Compared with the original YOLOv11n and mainstream comparative models such as YOLOv10n and YOLOv8n, DF-YOLOv11 achieves a better balance between detection accuracy, model complexity, and inference efficiency, exhibiting stronger robustness especially for small, tilted, and densely occluded targets. It provides a feasible technical solution for deployment of resource-constrained field mobile devices.