基于GA-DETR的自然场景谷类作物穗部检测方法

Method for Spike Detection of Cereal Crops in Natural Scenes Based on GA-DETR

摘要:

全球三大谷类作物（水稻、小麦、玉米）穗部的检测，是精准农业与谷类作物表型分析中的基础任务。然而，在复杂田间环境中，穗部分布密集、尺度差异显著且存在大量小目标严重影响检测精度。为解决这些问题，本研究提出一种基于RT-DETR架构的门控注意力DETR（GA-DETR），并引入3个创新组件：针对谷物穗部细节特征，设计门控机制C2F（GMC2F）模块，通过动态通道加权与跨阶段局部特征融合，提升骨干网络的特征判别能力。为解决谷物穗部形态差异导致的尺度不匹配问题，提出注意力上采样尺度序列特征融合（AUSSFF）模块，借助3D卷积强化多尺度特征依赖。针对无人机图像中小目标检测难题，提出FPIoU损失函数，结合目标尺寸自适应加权与难度感知分层策略，优化对难检测样本的处理能力。在水稻穗数据集（RiceR）、小麦穗数据集（GWHD）和无人机玉米雄穗数据集（MTC-UAV）上，GA-DETR性能优于基准模型RT-DETR及其他5种主流检测模型，mAP@0.5分别达到92.8%、91.7%和91.3%，对于水稻穗数据集（RiceR）模型内存占用量减少32.5%，浮点运算量降低14.4%。在GWHD数据集的穗部计数任务中，该框架性能超过5种主流框架，平均绝对误差（MAE）为5.650，均方根误差（RMSE）为7.383。GA-DETR有效平衡了检测精度与效率，为谷物穗部通用检测框架提供了跨物种特征建模范式，且兼容小麦、水稻、玉米等多种谷类作物，以及地面相机、无人机等不同采集平台的数据，可支持谷物自动化高通量田间表型监测，进一步推动精准农业发展。

Abstract:

The detection of inflorescences from the world’s three major cereal crops (rice panicles, wheat spikes, and maize tassels) is a fundamental task in precision farming and cereal crop phenotyping. However, accurate detection remains challenging due to dense distributions, significant scale variations, and small-target in complex environments, which substantially compromise the precision of detection. To tackle these issues, gated attention-DETR (GA-DETR), an architecture based on RT-DETR was proposed, which introduced three novel components: aiming at the delicate tip features of cereal inflorescences, a gated mechanism C2F (GMC2F) module was proposed to enhance backbone feature discrimination through dynamic channel weighting and cross-stage local feature integration. To address the scale mismatch caused by differences in the shapes of cereal inflorescences, an attention upsample scale sequence feature fusion (AUSSFF) module was proposed, which enhanced the robustness of multi-scale dependency modeling through 3D convolutions. For difficult small target in UAV images, a FPIoU loss function was proposed, which combined target-size adaptive weighting and difficulty-aware stratification to optimize the performance on hard samples. GA-DETR performed better than the baseline RT-DETR and five mainstream detection models on the RiceR dataset, GWHD dataset, and MTC-UAV dataset, including rice panicles, wheat spikes, and maize tassels, achieving mAP@0.5 of 92.8%, 91.7%, and 91.3%, respectively, while for RiceR dataset reducing model parameters by 32.5% and floating-point computational load by 14.4%. The proposed framework surpassed five state-of-the-art frameworks in inflorescence counting on GWHD dataset, achieving an MAE of 5.650 and an RMSE of 7.383. It effectively balanced accuracy and efficiency, providing a cross-species feature modeling paradigm for the universal detection framework of cereal inflorescence morphology. Compatible with diverse cereal crops (e.g., wheat, rice, maize) and data from different acquisition platforms (ground cameras, UAVs), it supported automated high-throughput field phenotyping monitoring of cereals, further advancing precision agriculture.