基于改进YOLO 11n‑seg的轻量化柑橘树冠层实例分割方法与喷雾试验

Lightweight Instance Segmentation Method and Spraying Experiment for Citrus Tree Canopy Based on Improved YOLO 11n‑seg

摘要:

针对果园精准施药中柑橘树冠层实例分割存在的实时性不足、复杂场景泛化能力差的技术问题，以及模型轻量化技术挑战，本文提出基于改进YOLO 11n?seg的轻量化实例分割模型。采用深度可分离卷积压缩关键层参数量至原结构11.3%，通过通道分离策略降低计算冗余;引入全局注意力机制实现通道?空间双路三维权重融合，有效抑制42.7%过曝区域误检;设计轻量化分割检测头融合多尺度特征，结合动态通道剪枝策略，在维持精度的同时降低31.4%计算量;通过深度阈值滤波及五维对抗性增强扩展至2 500幅样本，覆盖复杂柑橘园场景，构建了柑橘冠层RGB?D数据集。在枝叶遮挡35%及车速0.5 m/s的移动喷雾平台验证试验结果表明：改进模型分割精度（(Seg) AP50）达92.6%，较基准模型提升2.4个百分点，推理时间为0.178 s，优于实时性基准模型YOLOACT的12.7%;参数量仅2.53×10^6（为Mask R?CNN的24%）。基于改进模型的移动喷雾控制系统采用关键帧点云融合技术，将图像处理延时控制在320 ms，实现总延时404.93 ms下的位移偏差0.2025 m（车速0.5 m/s）。变量喷雾验证显示农药节省率达45.75%，药液分布变异系数降至10.87%，冠层中重叠区域无效喷施现象减轻。

Abstract:

Addressing critical technical bottlenecks in precision orchard spraying, specifically the insufficient real?time performance, poor generalization in complex scenarios, and model lightweighting challenges in citrus tree canopy instance segmentation, an innovative lightweight instance segmentation model was proposed based on an enhanced YOLO 11n?seg architecture. Key technical innovations included employing depthwise separable convolution (DSConv) to compress parameters in critical layers to merely 11.3% of the original structure, coupled with a channel separation strategy that significantly reduced computational redundancy, introducing a novel global attention mechanism (GAM) that achieved fused three?dimensional channel?spatial weighting through dimensional permutation operations, effectively suppressing 42.7% of overexposed region misdetections while enhancing edge feature representation and designing a lightweight segmentation detection head (LSDH) that integrated multi?scale feature fusion with dynamic channel pruning, reducing computational load by 31.4% while maintaining segmentation accuracy. To address data scarcity, a specialized RGB?D citrus canopy dataset containing 2 500 annotated samples captured by using Kinect DK depth cameras was constructed. This dataset was expanded through depth threshold filtering and five?dimensional adversarial augmentation (incorporating geometric transformations, photometric variations, and synthetic noise injection) to comprehensively represent complex orchard environments. Experimental validation under realistic 35% foliage occlusion conditions on a mobile spray platform operating at 0.5 m/s demonstrated the model?s superior performance: segmentation accuracy ((Seg) AP50) reached 92.6% (2.4 percentage points improvement over baseline), inference time achieved 0.178 s (12.7% faster than that of YOLOACT), and parameter count was reduced to only 2.53×10^6 (24% of Mask R?CNN). Field deployment results confirmed the system?s practical viability: utilizing keyframe point cloud fusion technology, image processing latency was constrained to 320 ms, enabling precise spraying control with just 0.2025 m displacement error at 0.5 m/s vehicle speed (total system delay: 404.93 ms). Variable?rate spraying validation showed 45.75% pesticide reduction, spray distribution uniformity (coefficient of variation) of 10.87%, and 17.53% reduction in overspray within overlapping canopy regions demonstrating improvements over conventional spraying methods.