Super-resolution and multi-scale fusion target detection algorithm based on improved YOLOv5
姚珊珊 1王静宇 1郝斌 1张飞 1高鹭 1任晓颖1
扫码查看
点击上方二维码区域，可以放大扫码查看
作者信息
1. 内蒙古科技大学信息工程学院,内蒙古包头 014000
折叠
摘要
为了提升目标检测算法在多尺度学习方面的能力,尤其是对小目标的检测能力,本文提出了一种基于改进YOLOv5的超分辨率和多尺度融合目标检测算法.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,并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的双向融合,将原YOLOv5算法的3尺度预测升级为4尺度预测,以此提高模型多尺度特征学习能力和对小目标的检测效果.实验结果表明,与YOLOv5s相比,改进后的模型在PASCAL VOC数据集中,mAP@0.5提高了2.8个百分点,mAP@0.5∶0.95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@0.5:0.95提高了5.2个百分点.改进后的YOLOv5模型在多尺度检测,尤其是小目标的检测效果方面得到了提升,并具有一定的应用价值.
Abstract
To enhance the multi-scale learning capacity of target detection algorithms,particularly for small targets,this paper proposes a super-resolution and multi-scale fusion target detection algorithm based on an improved YOLOv5 framework.Firstly,instead of the up-sampling operation of the original YOLOv5 model,the algorithm utilizes sub-pixel convolution to enhance the image resolution and pre-serve the information of small targets to the greatest extent possible.Secondly,the algorithm utilizes the parallel fast multi-scale fusion(PFMF)module to achieve two-way fusion of deep and shallow features.This upgrade from the original YOLOv5 algorithm's 3-scale prediction to 4-scale prediction improves the model's ability to learn multi-scale features and detect small targets.The experimental results demon-strate that compared with YOLOv5s,the improved model achieves a 2.8％and 3.5％increase in mAP@0.5 and mAP@0.5∶0.95,respectively,on the PASCAL VOC dataset.Similarly,on the MS COCO dataset,the improved model achieves a 4.3％and 5.2％increase in mAP@0.5 and mAP@0.5∶0.95,respectively.The experiments demonstrate the improved YOLOv5 model's enhanced capability in multi-scale detection,particularly for small targets,and indicate its potential practical value.
维普
万方数据