融合注意力机制和轻量化的目标检测方法研究  

Research on Fusion of Attention Mechanism and Lightweight Object Detection Method

作　　者：姚庆安[1] 孙旭[1] 冯云丛 乔石丽 YAO Qing-an;SUN Xu;FENG Yun-cong;QIAO Shi-li(College of Computer Science and Engineering,Changchun University of Technology,Changchun Jilin 130102,China)

机构地区：[1]长春工业大学计算机科学与工程学院,吉林长春130102

出　　处：《计算机仿真》2025年第2期252-258,525,共8页Computer Simulation

基　　金：吉林省自然科学基金项目(YDZJ202201ZYTS422);吉林省科技厅青年成长科技计划项目(20210508039RQ)。

摘　　要：针对YOLOv5s目标检测过程中小目标漏检导致整体检测精度下降的问题,提出一种基于注意力机制和轻量化的优化YOLOv5s目标检测算法。上述优化方法主要是在YOLOv5s的骨干网络中引入深度可分离卷积,在不降低检测精度的同时减少参数量,同时在颈部网络中采用CA注意力机制和Ghost卷积的搭配方式,在不增加参数量的同时有效的提升检测精度。仿真结果证明,优化后的YOLOv5s算法在PASCAL VOC2007数据集和RSOD数据集上的mAP值分别为72.8%和95.5%,相较于原YOLOv5s算法,mAP值分别提高了4.1%和2.0%,同时参数量也都减少了17%。通过仿真证明优化的YOLOv5s目标检测算法可以在提升检测精度的同时兼顾模型的轻量化。Aiming at the problem that the overall detection accuracy decreases due to the omission of small and medium targets during YOLOv5s target detection,an optimized YOLOv5s target detection algorithm based on attention mechanism and lightweight was proposed.This optimization method mainly introduced deep separable convolution into the backbone network of YOLOv5s to reduce the number of parameters without reducing the detection accuracy.Meanwhile,the collocation of CA attention mechanism and Ghost convolution was adopted in the neck network to effectively improve the detection accuracy without increasing the number of parameters.The simulation results show that the mAP values of the optimized YOLOv5s algorithm on PASCAL VOC2007 data set and RSOD data set are 72.8%and 95.5%,respectively.Compared with the original YOLOv5s algorithm,the mAP values are increased by 4.1%and 2.0%,respectively,and the number of parameters is reduced by 17%.The simulation experiments show that the optimized YOLOv5s target detection algorithm can not only improve the detection accuracy,but also take into account the lightweight of the model.

关 键 词：目标检测 注意力机制 轻量化 仿真 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]