改进YOLOv3的轻量化漂浮物检测算法  

Lightweight floating object detection algorithm based on improved YOLOv3

作　　者：任英杰 李传奇[1] 王薇[1] 葛召华 REN Yingjie;LI Chuanqi;WANG Wei;GE Zhaohua(School of Civil Engineering,Shandong University,Jinan 250061,Shandong,China;Shandong Water Conservancy Comprehensive Service Center,Jinan 250013,Shandong,China)

机构地区：[1]山东大学土建与水利学院,山东济南250061 [2]山东省水利综合事业服务中心,山东济南250013

出　　处：《水利水电技术（中英文）》2023年第10期170-179,共10页Water Resources and Hydropower Engineering

基　　金：山东省自然科学基金(ZR2021ME030);深圳市可持续发展科技专项项目(KCXFZ20201221173407021);济南市水务科技项目(JNSWKJ202106)。

摘　　要：【目的】为解决水域监控下漂浮物检测效率低、检测模型复杂度高的问题,提出一种基于改进YOLOv3的轻量化漂浮物检测算法。【方法】使用轻量级网络MobileNetv3代替YOLOv3的主干特征提取网络Darknet53以降低模型计算量和参数;构建简化版加权双向特征金字塔网络(Bi-FPN-tiny)以进行多尺度特征的加权融合;利用Focal Loss优化损失函数,加强对于困难样本的学习。为验证所提算法的有效性,建立了PASCAL VOC格式的漂浮物数据集,并进行数据标注和增广。【结果】结果表明:改进后的算法平均精度均值(mAP)达到92.8%,比原算法提高了7.1%;在NVIDIA Quadro P2200显卡下检测速度达到了86 fps/s,高于YOLOv3算法的47 fps/s;模型体积为43.7 MB,仅为初始算法的17.7%。【结论】改进YOLOv3是一种性能优越且轻量化的模型,为在移动端进行实时漂浮物检测提供了新的契机。[Objective]To address the challenges of low detection efficiency and high model complexity in detecting floating objects under water monitoring,a lightweight floating object detection algorithm based on improved YOLOv3 is proposed.[Methods]The proposed algorithm employs MobileNetv3 as the feature extraction network and constructs a simplified version of weighted bidirectional feature pyramid(Bi-FPN-tiny) for feature fusion.Focal Loss is used to optimize the loss function and strengthen the learning of difficult samples.To evaluate the effectiveness of the improved algorithm,a floating object dataset in PASCAL VOC format is established and tested after data annotation and amplification.[Results]Experimental result show that the mean average accuracy(mAP) of the improved algorithm reaches 92.8%,which is 7.1% higher than the original algorithm.The detection speed of NVIDIA Quadro P2200 is 86 fps/s,higher than the 47 fps/s of YOLOv3.The model size is 43.7 MB,which is only 17.7% of the initial algorithm.[Conclusion]The improved YOLOv3 is a high-performance and lightweight model,providing new opportunities for real-time floating object detection on mobile devices.

关 键 词：YOLOv3算法 漂浮物 目标检测 轻量化 特征融合 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]