To address the detection difficulty of small-diameter tree trunks in complex forest stands—characterized by slender shapes, frequent occlusion, and high similarity to background textures—this study proposes a lightweight improved model, termed LDM-YOLO, built upon YOLO11n using a self-constructed small-diameter tree image dataset. The acronym LDM is derived from LSKA (large separable kernel attention), DySample dynamic upsampling, and MBConv-Detect. Specifically, a C3k2-LSKA module (cross-stage partial network bottleneck with two convolutions integrated with large separable kernel attention) is introduced into the Backbone to enhance the representation of slender trunks and their contextual features. In the Neck, DySample dynamic upsampling is adopted to refine multi-scale feature alignment. Moreover, an MBConv-Detect structure (mobile inverted bottleneck convolution–detection head) is applied to the regression branch of the detection head to reduce parameter count and computational cost. Experimental results on the self-built dataset show that LDM-YOLO improves mAP50 and mAP50–95 by 3.1 and 2.3 percentage points, respectively, compared with YOLO11n, while reducing GFLOPs from 6.3 G to 5.1 G, thereby achieving higher detection accuracy under low inference overhead. Compared with representative methods such as Faster R-CNN, YOLOv5n, YOLOv8n, and YOLOv10n, LDM-YOLO delivers superior overall detection performance at comparable or lower computational complexity. The results demonstrate that the proposed model better balances accuracy and lightweight design for small-diameter tree detection, providing a useful reference for subsequent forestry intelligent equipment applications on resource-constrained platforms.
以卷积神经网络(convolutional neural network,CNN)为核心的深度学习模型,能够从海量图像数据中自动学习和提取高维、鲁棒性的特征,摆脱了传统图像处理方法对人工设计特征的依赖。单阶段目标检测算法YOLO系列[3]凭借端到端高速检测和较高准确率,已成为林业树干检测的主流选择。例如,Badgujar等[4]采用YOLOv5对天然牧场环境中的东红杉树干进行检测,经过迁移学习训练的模型在测试中达到约87.8%的精度、84.3%的召回率和88.9%的平均精度,能够识别不同粗细的树干并有效区分树干与树枝。在密林场景中,Da Silva等[5]比较了多种算法,结果表明YOLOv7及其变体YOLOR在大规模林地树干数据集上取得了最高约90%的F1分数,而且在图形处理器(graphics processing unit,GPU)上仍可实现毫秒级的实时检测。针对更复杂的应用需求,Liu等[6]在YOLOv7中融入注意力机制并改进损失函数,用于油茶林收获机器人视觉中的树干识别,使模型平均精度提升至89.2%、召回率提高到94%,检测性能显著优于原始YOLOv7模型。Zhang等[7]在YOLOv8基础上引入金字塔特征增强网络、Transformer模块和坐标注意力机制,提出了PCC-YOLO金字塔增强网络、上下文Transformer模块、坐标-通道融合注意力机制YOLO(a YOLOv8-based model integrating PENet,CoT-Net,and Coord-SE attention,PCC-YOLO)树干检测模型,在果园树干数据集上实现了82.6%的mAP和约143帧/s的检测速度,表现出卓越的精度与效率平衡。Zhang等[8]提出了面向林地场景的轻量级YOLOv11-TrunkLight模型,引入新型高效骨干网络和可变形注意力模块,有效提升了树干目标的检测性能。杨灿等[9]提出空间感知与多尺度特征增强自适应网络YOLO(spatial‑aware and FastScaleNet YOLO,SSFYOLO),通过空间感知模块(spatial-aware module,SAM)、多尺度特征增强网络FastScaleNet与加权IoU机制,显著增强了模型在复杂林地小目标检测中的精度与稳定性。胡峻峰等[10]针对森林移动机器人算力受限与光照变化大的问题,提出基于YOLOv7-tiny的轻量级检测方法,引入部分通道卷积高效层聚合网络(partial efficient layer aggregation network,P-ELAN)与α‑完整交并比(alpha-complete Intersection over union,alpha-CioU),提升推理速度同时保持检测精度,适配边缘设备部署。在农业无人机低空作业场景下,徐相华等[11]提出双坐标注意力-单阶段目标检测模型(dual coordinate attention you only look once,DCA-YOLO),引入动态通道注意力与轻量化主干网络,实现复杂背景下对病虫害树木的高效精准识别。此外,针对与树干结构高度关联的树木年轮信息提取任务,王品博等[12]提出基于YOLOv8的DCW-YOLOv8)可扩张残差注意力模块-轻量级通用上采样算子-动态非单调聚焦机制边界框损失函数的YOLOv8模型(dilation-wise residual-content-aware reassembly of features-weighted interpolation of sequential evidence for intersection over union,DCW-YOLOv8),实现了年轮实例分割与宽度测量,为林木生长特征自动化识别提供了新思路。
BAIY, YANGH, WENJ,et al.Study on forest structure diversity based on the neighbourhood trees[J].Journal of Beijing Forestry University,2020,42(6):52-58.
[3]
ZHAOZ Q, ZHENGP, XUS,et al.Object detection with deep learning:A review[J].IEEE Transactions on Neural Networks and Learning Systems,2019,30(11):3212-3232.
[4]
REDMONJ, DIVVALAS, GIRSHICKR,et al.You only look once:Unified,real-time object detection[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition.June 27-30,2016,Las Vegas,NV,USA,IEEE,2016:779-788.
[5]
BADGUJARC, FLIPPOD, GUNTURUS,et al.Tree trunk detection of eastern red cedar in rangeland environment with deep learning technique[J].Croatian Journal of Forest Engineering,2023,44(2):357-368.
[6]
SILVA D QDA, DOS SANTOSF N, FILIPEV,et al.Edge AI-based tree trunk detection for forestry monitoring robotics[J].Robotics,2022,11(6):136.
[7]
LIUY, WANGH R, LIUY H,et al.A trunk detection method for Camellia oleifera fruit harvesting robot based on improved YOLOv7[J].Forests,2023,14(7):1453.
[8]
ZHANGS, SUNY, QIAOY,et al.PCC-YOLO:A trunk detection algorithm for orchard robots based on improved YOLOv8[J].Agriculture,2025,15(16):1786.
[9]
ZHANGS Q, ZHENGY B, BIR G,et al.Trunk detection in complex forest environments using a lightweight YOLOv11-TrunkLight algorithm[J].Sensors,2025,25(19):6170.
XUX H, ZHOUD J, YUC R,et al.Improved model for low altitude detection of trees infected by pests and diseases using agricultural drones based on DCA-YOLO[J].Transactions of the Chinese Society for Agricultural Machinery,2025,56(10):479-491.
WANGP B, WANGY C, WANGS Y,et al.Study on tree ring instance segmentation and information detection based on improved YOLOv8[J].Chinese Journal of Wood Science and Technology,2025,39(3):7-18.
[18]
LAUK W, POL M, REHMANY A U.Large separable kernel attention:rethinking the large kernel attention design in CNN[J].Expert Systems With Applications,2024,236:121352.
[19]
AZADR, NIGGEMEIERL, HUTTEMANNM,et al.Beyond self-attention:Deformable large kernel attention for medical image segmentation[C]//2024 IEEE/CVF Winter Conference on Applications of Computer Vision.2024:1287-1297.
[20]
LIUW Z, LUH, FUH T,et al.Learning to upsample by learning to sample[C]//2023 IEEE/CVF International Conference on Computer Vision.2023:6027-6037.
[21]
TANM X, LEQ V.EfficientNetV2:Smaller models and faster training[C]//2021 International Conference on Machine Learning.2021:10096-10106.
[22]
RENS Q, HEK M, GIRSHICKR,et al.Faster R-CNN:towards real-time object detection with region proposal networks[C]//IEEE Transactions on Pattern Analysis and Machine Intelligence.IEEE,2015,28:1137-1149.