基于多尺度特征融合的树木检测算法 DDC-YOLO

打开文本图片集

中图分类号：S758 文献标识码：A DOI：10.7525/j.issn.1006-8023.2026.01.01

TreeDetection Algorithm Based on Multi-scale FeatureFusion -DDC-YOLO

ZHANG Zhijie，WANGQin*

Abstract： This paper proposes a detection algorithm DDC-YOLO based on improved YOLOv10 to address the issues oclusion interference insufficient lighting in tree detection.Firstly，a dynamic convolutionalmix block（DCMB） wasdesigned to enhance the multi-scale feature fusioncapabilitythrough adaptive dynamic convolution，solving the problem singularity in traditional convolutionkernels；Secondly，adual backbonedynamic feature fusion network was proposed，combining the backbone structures RT-DETR YOLOv10， utilizing the dynamic alignment fusion （DAF）module to adjust feature weightsenhance the model's adaptability to diferent features；Further introduced pyramid context feature extraction spatialfeature reconstruction techniques tooptimize the neck network achieve deep fusion multi-levelsemantic information.The experimentwas validated basedontheself built dataset TreeImages （containing 7475 images）， the results showed that the mAP50 DDC-YOLO reached 46. 7% ，which was 5. 0 percentage points higherthan the original YOLOv10 model.The parameter size decreased from 2.27 M to 2.26 M（a decrease 0.44% ）， the detection speed（FPS）increased from 2O2 to 254（an increase 25.4% ）.The improved modelexhibits higherrobustnessreal-time performanceincomplex scenarios，providing an eficient solutionfor forest resource surveys.

Keywords：YOLOvlO；object detection；computer vision；dual-backbone dynamic fusion；multi-scale feature reconstruction

0 引言

在森林资源管理中，树木调查方法仍以传统人工测量和自测估算为主，这种方式耗时长且效率低。（剩余16510字）