基于DGA-DETR模型的芦笋选择性采收精准识别方法研究

Precise Identification Method for Selective Harvesting of Asparagus Based on DGA-DETR Model

摘要:

针对芦笋机器人采收作业过程中,芦笋植株分布密集、遮挡严重导致感知精度不足、漏检率高、重叠实例粘连及跨域泛化能力弱等问题,对芦笋细长几何形态特性和密集遮挡的空间特征进行深入分析,提出了一种融合动态门控注意力与边缘感知下采样机制的实例分割模型DGA-DETR(Dynamic gated attention-detection Transformer)。通过设计动态门控注意力(DGA)模块,利用实例语义驱动的动态门控机制实现跨空间与跨尺度的特征精准聚焦,从机理层面有效抑制密集遮挡下背景噪声干扰;构建了边缘感知上下采样机制,以强化边界解耦能力,提升模型对芦笋细长形态及遮挡交界区域的分割建模精度。为验证分割模型的有效性,构建了不同光照条件及遮挡程度的机收芦笋数据集用于模型训练,涵盖像素、场景、语义多层次域偏移,并搭建跨域数据集进行测试。实验结果表明,所提DGA-DETR模型在测试集上的掩码精确率、掩码召回率、Mask mAP@0.5和Mask mAP@0.5:0.95分别达到91.60%、67.24%、90.98%和59.13%,较基线模型分别提升5.33、8.43、6.56、9.51个百分点,显著降低了密集遮挡条件下芦笋实例漏分割与掩码不连续现象。跨域测试结果表明,该方法在遮挡与复杂背景条件下芦笋的精准识别表现出良好鲁棒性与泛化能力,研究结果可为其他嫩茎类作物在复杂环境下机器人采收作业提供技术参考。

Abstract:

Aiming at the technical challenges in the robotic harvesting process of asparagus, such as insufficient perception accuracy, high miss rate, adhesion of overlapping instances, and weak cross- domain generalization caused by dense plant distribution and severe occlusion, an in-depth analysis of the slender geometric morphology and spatial features of dense occlusion in asparagus was conducted. An instance segmentation model, dynamic gated attention-detection transformer (DGA-DETR), which integrated dynamic gated attention and edge awareness, was proposed. By designing a dynamic gated attention (DGA) module, an instance-semantic-driven dynamic gating mechanism was utilized to achieve precise feature focusing across space and scales, effectively suppressing background noise interference under dense occlusion at the mechanistic level. Additionally, an edge-aware up-and-down sampling mechanism was constructed to strengthen boundary decoupling capabilities, improving the model’s segmentation modeling accuracy for the slender morphology of asparagus and occlusion boundary regions. To verify the effectiveness of the segmentation model, an asparagus harvesting dataset covering different lighting conditions and occlusion levels was constructed for model training, encompassing multi-level domain shifts in pixels, scenes, and semantics, and a cross-domain dataset was established for testing. Experimental results demonstrated that the proposed DGA-DETR model achieved Mask precision, Mask recall, Mask mAP@ 0. 5, and Mask mAP@ 0. 5:0. 95 of 91. 60% , 67. 24% , 90. 98% , and 59. 13% , respectively, representing improvements of 5. 33, 8. 43, 6. 56, and 9. 51 percentage points over the baseline model, and significantly reducing missed instances and mask discontinuities under dense occlusion conditions. Cross-domain test results indicated that this method exhibited excellent robustness and generalization capabilities for the precise identification of asparagus under occlusion and complex backgrounds. The research results can provide a technical reference for the robotic harvesting of other tender-stem crops in complex environments.