2025年01期 v.46 29-35页

刘印1  龚长友2  徐国栋1

（1.西南林业大学，云南 昆明 650224

2.新疆生产建设兵团兴新职业技术学院，新疆 铁门关 841007）

摘要：针对目标检测算法部署在边缘设备的轻量化需求，提出一种基于改进YOLOv10的轻量化目标检测算法（CMD-YOLO算法）。该算法利用跨尺度特征融合模块对YOLOv10算法的网络结构进行改进，减少了算法模型的参数量与计算量；采用基于Mamba的线性注意力机制改进的部分自注意力模块替换传统的部分自注意力模块，进一步降低了算法模型的参数量；利用空间深度转换卷积模块替换部分传统卷积模块，增强了算法模型对下采样细节信息的提取能力；利用动态上采样器DySample替换传统的上采样模块，在保持上采样精度的同时，降低了算法模型的计算延迟。实验结果表明，CMD-YOLO算法与YOLOv10-n算法相比，在检测精度略微提升的同时，算法模型参数量降低了30.5%，计算量下降了19%，权重文件缩小了29.3%，计算延迟减少了8.8%，能够满足目标检测算法部署在边缘设备中的轻量化需求。

关键词：目标检测算法；YOLOv10算法；跨尺度特征融合模块；Mamba线性注意力机制；空间深度转换卷积模块；动态上采样器 

中图分类号：TP391.41           文献标志码：A          文章编号：1674-2605(2025)01-0004-07

DOI：10.3969/j.issn.1674-2605.2025.01.004                     开放获取

Lightweight Object Detection Algorithm Based on Improved YOLOv10

LIU Yin1  GONG Changyou2  XU Guodong1

(1.Southwest Forestry University, Kunming 650224, China

2.Bingtuan Xingxin Vocational and Technical College, Tiemenguan 841007, China)

Abstract: Aiming at the lightweight requirements of deploying object detection algorithms on edge devices, a lightweight object detection algorithm based on improved YOLOv10 (CMD-YOLO algorithm) is proposed. This algorithm utilizes a cross-scale feature fusion module to improve the network structure of YOLOv10 algorithm, reducing the parameter and computational complexity of the algorithm model; Adopting a Mamba based linear attention mechanism to improve the partial self attention module and replace the traditional partial self attention module, further reducing the parameter count of the algorithm model; Replacing some traditional convolution modules with spatial depth conversion convolution modules enhances the algorithm model's ability to extract downsampling detail information; By using the dynamic UpSampler DySample to replace the traditional upsampling module, the computational delay of the algorithm model is reduced while maintaining upsampling accuracy. The experimental results show that compared with the YOLOv10-n algorithm, the CMD-YOLO algorithm has slightly improved detection accuracy, reduced model parameters by 30.5%, decreased computational complexity by 19%, reduced weight files by 29.3%, and reduced computational latency by 8.8%, which can meet the lightweight requirements of object detection algorithm deployment in edge devices.

Keywords: object detection algorithm; YOLOv10 algorithm; cross-scale feature fusion module; Mamba-like linear attention mechanism; space to depth Conv module; dynamic UpSampler