基于改进YOLOv5的超分辨率和多尺度融合目标检测算法
Super-resolution and multi-scale fusion target detection algorithm based on improved YOLOv5
投稿时间:2022-12-19  修订日期:2023-03-30
DOI:
中文关键词:  目标检测  YOLOv5算法  子像素卷积  多尺度融合
英文关键词:object detection  YOLOv5 algorithm  sub-pixel convolution  multi-scale fusion  
基金项目:内蒙古自治区科技计划项目
作者单位邮编
姚珊珊 内蒙古科技大学 014010
王静宇 内蒙古科技大学 
郝斌* 内蒙古科技大学 014010
张飞 内蒙古科技大学 
高鹭 内蒙古科技大学 
任晓颖 内蒙古科技大学 
摘要点击次数: 149
全文下载次数: 0
中文摘要:
      为了使得目标检测算法在多尺度检测上进一步提升正确率,解决小目标物体分辨率较低、可用信息较少和细节信息不明显等问题,本文提出了一种改进的YOLOv5目标检测算法,实现多尺度目标自动检测。首先,该算法使用子像素卷积代替原YOLOv5模型的上采样操作,提高图像的分辨率,尽可能保留小目标的信息。其次,使用并行快速多尺度融合(Parallel Fast Multi-scale Fusion,PFMF)模块实现深层特征和浅层特征双向融合,以此提高对小目标的检查效果。除此以外,PFMF模块还将原YOLOv5算法的3尺度预测升级为4尺度预测,提高了多尺度特征的学习能力。实验结果表明,改进后的模型与YOLOv5s相比,在PASCAL VOC数据集中,mAP@0.5提高了2.8个百分点,mAP@.5:.95提高了3.5个百分点;在MS COCO数据集中,mAP@0.5提高了4.3个百分点,mAP@.5:.95提高了5.2个百分点。由此可见,改进后的YOLOv5模型在多尺度检测上的有效性。
英文摘要:
      In order to improve the accuracy of target detection algorithm in multi-scale detection and solve the problems of low resolution, less available information and unclear detail information of small target objects, an improved YOLOv5 target detection algorithm is proposed to realize multi-scale target automatic detection.Firstly, the sub-pixel convolution is used to replace the up-sampling operation of the original YOLOv5 model to improve the resolution of the image and preserve the information of small objects as much as possible.Secondly, the parallel fast multi-scale fusion (PFFF) module is used to realize the bidirectional fusion of deep features and shallow features, so as to improve the inspection effect of small targets.In addition, the PFFF module also upgrades the 3-scale prediction of the original YOLOv5 algorithm to 4-scale prediction, which improves the learning ability of multi-scale features.The experimental results show that, compared with YOLOv5s, the mAP@0.5 and mAP@.5:.95 of the improved model are improved by 2.8% and 3.5% respectively in PASCAL VOC dataset. In the MS COCO dataset, mAP@0.5 improved by 4.3 percentage points and mAP@.5:.95 improved by 5.2 percentage points.Thus, the improved YOLOv5 model is effective in multi-scale detection.
    下载PDF阅读器
关闭

版权所有:《光电子·激光》编辑部  津ICP备12008651号-1
主管单位:天津市教育委员会 主办单位:天津理工大学 地址:中国天津市西青区宾水西道391号
技术支持:北京勤云科技发展有限公司