摘要
视频帧中目标模糊、遮挡和形变是影响视频目标检测精度的重要原因,针对此类问题,提出一种DG-YOLOv8n算法。首先,基于可变形卷积对骨干网络中的C2f模块进行了重新设计,以增强骨干网络对于目标变化的处理能力;其次,在颈部网络引入GAM全局注意机制,放大全局交互表示来提高算法的性能;最后,在ImageNet VID数据集上的实验结果表明,改进的DG-YOLOv8n算法的平均精度为84.5%,较原YOLOv8n算法提高了6.1个百分点,验证了改进算法的有效性。
The blurring,occlusion,and deformation of targets in video frames are important factors affecting the accuracy of video object detection.To address these issues,a DG-YOLOv8n algorithm is proposed.Firstly,the C2f module in the backbone net-work was redesigned based on deformable convolution to enhance its ability to handle target changes.Secondly,the GAM global at-tention mechanism is introduced into the neck network to amplify the global interactive representation and improve the performance of the algorithm.Finally,the experimental results on the ImageNet VID dataset showed that the improved DG-YOLOv8n algorithm had an average precision of 84.5%,which was 6.1 percentage point higher than the original YOLOv8n algorithm,verifying the effec-tiveness of the improved algorithm.
作者
魏一帆
郭本华
粟长权
钱淑渠
Wei Yifan;Guo Benhua;Su Changquan;Qian Shuqu(School of Information,Guizhou University of Finance and Economics,Guiyang 550025,China;College of Mathematics and Computer Science,Anshun University,Anshun 561000,China)
出处
《现代计算机》
2024年第21期20-25,共6页
Modern Computer
基金
贵州省教育厅创新群体重大资助项目(黔教合KY字[2018]034)
安顺学院研究生创新基金资助项目(asxyyjscx202401)。
关键词
目标检测
视频目标检测
可变形卷积
GAM全局注意机制
YOLOv8
object detection
video object detection
deformable convolution
global attention mechanism
feature aggregation
YOLOv8
作者简介
魏一帆(1998-),女,河南安阳人,硕士研究生,研究方向为深度学习、目标检测;通信作者:郭本华(1964-),男,贵州兴义人,博士,硕士生导师,教授,研究方向为复杂网络、深度学习,E-mail:gbhua@126.com;粟长权(1996-),男,贵州黔东南人,硕士研究生,研究方向为深度学习、医学图像分割;钱淑渠(1978-),男,安徽枞阳人,博士,硕士生导师,教授,研究方向为智能优化算法。
