改进YOLOv7的小目标检测算法研究
Research on improved YOLOv7 algorithm for small target detection in aerial photography
朱琳 1代涛 1黎青松1
作者信息
- 1. 西华大学 汽车与交通学院,四川 成都 611730
- 折叠
摘要
针对小目标检测时因目标像素小、分布密集且相互遮挡等原因导致的检测精度低、漏检率高等问题,提出一种基于YOLOv7 的改进算法.使用ACON激活函数更换SiLU激活函数,增强算法的拟合能力;使用SPPFCSPC替换算法中的SPPCSPC结构,结合BiFormer注意力机制改进SPPFCSPC,使之提高模型检测效果的同时加快模型的速度;使用SPD-Conv卷积改进算法中的MP结构,提高模型的特征提取能力;在WIoU中引入辅助框设计Inner-WIoU,并结合NWD损失函数改进YOLOv7 的损失函数.经过在VisDrone2019 数据集上的实验证明:在没有增加太多参数量的前提下,改进后的算法比原算法的mAP@0.5 提高 4.39%,验证了改进算法对提高模型性能的有效性.
Abstract
A YOLOv7 based improved algorithm was proposed to address the issues of low detection accuracy and high missed detection rate in small object detection due to small target pixels,dense distribution,and mutual occlusion.Firstly,replace the SiLU activation function with the ACON activation function to enhance the algorithm's fitting ability;replace the SPPCSPC structure in the algorithm with SPPFCSPC,and improve SPPFCSPC by combining BiFormer attention mechanism to improve model detection performance and accelerate model speed;using SPD-Conv convolution to improve the MP structure in the algorithm and enhance the feature extraction ability of the model;introduced auxiliary boxes in WIoU to design Inner-WIoU,and improved YOLOv7's loss function by combining NWD loss function.Through experiments on the VisDrone2019 dataset,it had been proven that the improved algorithm outperforms the original algorithm without adding too many parameters mAP@0.5 improved by 4.39%,verifying the effectiveness of the improved algorithm in improving model performance.
关键词
小目标检测/激活函数/注意力机制/特征提取/损失函数Key words
small target detection/activation function/attention mechanism/feature extraction/loss function引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据