目的:为解决目前无人机可见光系统检测小目标时准确率和实时性低的问题,提出一种基于改进YOLOv8的可见光小目标检测方法.方法:选取由主干网络(Backbone)、颈部模块(Neck)和头部模块(Head)组成的YOLOv8网络作为基础框架构建AGC-YOLO模型.首先,在Backbone部分融入卷积注意力模块(convolutional block attention module,CBAM),提高模型的特征表达能力;其次,将部分传统卷积模块替换为可改变核卷积模块AKconv,减少网络参数量;最后,在Neck部分采用Gold-YOLO模块,提高对不同尺寸目标的检测能力.选用VisDrone2019数据集分别进行消融实验和对比实验,通过平均精度均值(mean average precision,mAP)、每秒传输帧数(frames per second,FPS)、每秒 10 亿次的浮点运算数(giga floating-point operations per second,GFLOPs)和参数量(parameters)评估 AGC-YOLO模型对小目标检测的效果.结果:AGC-YOLO模型的FPS为31.90,GFLOPs和Parameters分别为9.20和11.30 M,达到无人机实时性的检测速度要求(FPS不低于30).虽然AGC-YOLO模型的GFLOPs和Parameters比轻量化模型YOLOv8n、Ghost-YOLO和YOLO-BiFPN有所增加,但是mAP50(mAP50表示在交并比为0.5时的mAP)分别提高了15％、6％和5％.结论:提出的方法在提高检测速度、减少参数量、保障检测精度方面表现良好,在无人机可见光小目标检测方面具有良好的应用前景.
Research on UAV visible light small target detection method based on improved YOLOv8
Objective To propose an improved Y OLOv8-based visible small target detection method to solve the problems of the UAV visible light system in accuracy and timeliness when applied to measuring small targets.Methods A YOLOv8 network consisting of Backbone,Neck and Head was used as the base framework to construct an AGC-YOLO model.Firstly,a convolutional block attention module(CBAM)was incorporated into Backbone to improve the feature expression of the model;secondly,some traditional convolution modules were replaced with the changeable kernel convolution module AKconv to reduce the network parameters;finally,a Gold-YOLO module was involved in Neck to enhance the detection ability for targets with different sizes.VisDrone2019 dataset was used to carry out ablation and comparison experiments,and the efficacy of the AGC-YOLO model for detecting small targets was evaluated in terms of mean average precision(mAP),frames per second(FPS),giga floating-point operations per second(GFLOPs)and parameters.Results The AGC-YOLO model had the FPS,GFLOPs and parameters being 31.90,9.20 and 11.30 M respectively,meeting the real-time detection speed requirements of drones(FPS not lower than 30),in which the mAP50(the mAP with the intersection over union being 0.5)was increased by 15％,6％and 5％when compared with those of the lightweight YOLOv8n,Ghost-YOLO and YOLO-BiFPN models.Conclusion The method proposed behaves well in speed,decreased parameters and precision,and is worthy promoting for UAV visible small target detection.[Chinese Medical Equipment Journal,2025,46(1):1-6]
万方数据
维普
