Dense Small Target Image Detection Algorithm Based on the Improved YOLOv8  

基于改进YOLOv8密集小目标图像检测算法

作　　者：MA Jing-yu SI Zhan-jun 马静宇;司占军(天津科技大学人工智能学院,天津300457)

机构地区：[1]College of Artificial Intelligence,Tianjin University of Science and Technology,Tianjin 300457,China

出　　处：《印刷与数字媒体技术研究》2025年第2期65-71,97,共8页Printing and Digital Media Technology Study

摘　　要：The YOLOv8 model faces challenges with dense target distribution and small size,resulting in lower accuracy in dense small target detection.To address these issues,an improved small target detection algorithm based on the YOLOv8 model was proposed in this paper.Firstly,the Global Attention Module(GAM)was introduced to enhance data prediction capability and model expression ability.Secondly,the Space-to-Depth(SPD)module was incorporated into the backbone network for fine-grained feature information learning to mitigate feature information loss due to down-sampling.Finally,a 160 pixels×160 pixels feature layer was added to expand small target feature information and effectively reduce instances of missed targets.Experimental validation on the public VisDrone2019 UAV small target detaset demonstrated that the proposed model achieves significant performance improvement in small target detection tasks compared to existing models,exhibiting higher accuracy.YOLOv8模型在检测密集小目标时会产生目标分布密集、尺寸过小导致检测精度低的问题。为了解决这些问题,本研究提出了一种基于改进YOLOv8模型的小目标检测方法。首先,通过引入全局注意力机制(Global Attention Module,GAM),加强了对数据的预测能力,增强了模型的表达能力。其次,在Backbone网络中引入空间深度模块(Space-to-Depth,SPD)用于学习细粒度的特征信息,以减少下采样过程中特征信息的损失。最后,增加了160像素×160像素的特征图层,扩展小目标的特征信息,有效减少了目标漏检的现象。改进后的算法在公开的VisDrone2019无人机小目标数据集上进行了实验,并与现有算法进行对比,实验结果表明,本研究改进模型在小目标检测任务中取得了显著的性能提升,具有更高的准确性和鲁棒性。

关 键 词：YOLOv8 Small targets GAM SPD 

分 类 号：TS8[轻工技术与工程] TP39[自动化与计算机技术—计算机应用技术]