摘要
为了解决无人机视角下场景复杂多变、目标尺度变化大且大多数为小目标等问题,提出一种改进的YOLOv8算法。首先,将传统的FPN替换为BiFPN,通过其双向信息流和加权特征融合机制,能够更有效地整合来自网络不同层级的信息;其次,将检测头替换为Dynamic Head,通过动态融合机制适应不同尺度的特征,从而提升对各种尺寸目标的检测能力;最后,引入Soft-NMS,通过降低重叠框的置信度,更有效地区分和保留紧密排列的目标,进而提升模型对密集型小目标的检测性能。结果表明,YOLOv8-BDS算法在VisDrone2019数据集上的平均精度值(mAP@IoU=0.5)达到了54.2%,相比原基准网络提升了6.8个百分点。该算法能够更好地满足密集型小目标的检测需求。
In order to solve the problems of complex and variable scenes from the perspective of drones,large changes in target scale from their perspective,and mostly small targets,an improved YOLOv8 algorithm is proposed.Firstly,changing the traditional FPN to BiFPN,through its bidirectional information flow and weighted feature fusion mechanism,can more effectively integrate information from different levels of the network;Secondly,the detection head is replaced with Dynamic Head,which adapts to features of different scales through dynamic fusion mechanism,improving the detection ability for targets of various sizes;Finally,by introducing Soft-NMS and reducing the confidence of overlapping boxes,it is possible to more effectively distinguish and preserve these tightly arranged targets,thereby improving the detection performance of the model for dense small targets.The results indicate that the YOLOv8-BDS algorithm has an average accuracy value on the VisDrone2019 dataset(mAP@IoU=0.5)achieved 54.2%,which is 6.8 percentage points higher than the original benchmark network.This algorithm can better meet the requirements of dense small object detection.
作者
张弛
游浩
高谦
张光明
谈栋华
彭菊红
ZHANG Chi;YOU Hao;GAO Qian;ZHANG Guangming;TAN Donghua;PENG Juhong(School of Artificial Intelligence,Hubei University,Wuhan 430062,China)
出处
《软件导刊》
2025年第8期196-200,共5页
Software Guide
作者简介
张弛(2000-),男,湖北大学人工智能学院硕士研究生,研究方向为人工智能算法设计;游浩(1999-),男,湖北大学人工智能学院硕士研究生,研究方向为计算机视觉;高谦(2001-),女,湖北大学人工智能学院硕士研究生,研究方向为人工智能;张光明(1999-),男,湖北大学人工智能学院硕士研究生,研究方向为自然语言处理;谈栋华(2001-),男,湖北大学人工智能学院硕士研究生,研究方向为语义分割;通讯作者:彭菊红(1978-),女,博士,湖北大学人工智能学院副教授,研究方向为信号处理、电路建模与分析、人工智能算法。
