基于YOLO优化的轻量级目标检测网络  

Light-weight Object Detection Network Optimized Based on YOLO Family

作　　者：许虞俊 李晨[1] XU Yu-jun;LI Chen(School of Electronic Science and Engineering,Southeast University,Nanjing 210096,China)

机构地区：[1]东南大学电子科学与工程学院,南京210096

出　　处：《计算机科学》2021年第S02期265-269,共5页Computer Science

摘　　要：目标检测是计算机视觉领域中一个相当活跃的研究领域,通过设计大型的深度卷积神经网络来提高目标检测的精度是一种十分有效的方法,然而目前在内存受限的应用场景中并不支持部署大型目标检测网。针对以上问题,文中提出了一种基于You Only Look Once(YOLO)系列单镜头目标检测网络设计原则的轻量级目标检测网,融合了GhostNet中的Ghost Module模块,并参考了MobileNet-v3中的通道注意力模块SE(Squeeze-and-Excitation),在卷积块中加入更优的ECA(Efficient Channel Attention)模块可以更好地利用可用的网络容量,使得网络在减少体系结构和计算的复杂度以及提升模型性能之间实现强的平衡;并且采用了Distance-IoU loss来解决检测框定位不准的问题,有效地提升了网络的收敛速度。最终模型的参数数量被压缩到了1.54 MB,小于YOLO Nano(即4.0MB),并且在VOC2007测试集上的mAP达到了72.1%,高于现有的YOLO Nano(即69.1%)。Object detection is an active research field in the computer vision field.It is a very effective method to improve object detection precision by designing a large-scale deep convolutional neural network.However,it is unfavorable to deploy a large-scale object detection network in memory-limited applications.To solve the above problems,this paper proposes a light-weight object detection network which is based on design principles from the YOLO family of single-shot object detection network architectures.This network integrates the Ghost Module in GhostNet,in addition,a better Efficient Channel Attention(ECA)module is added to the convolution block by referring to the Squeeze-and-Excitation(SE)module in MobileNet-v3.This module can make better use of the available network capacity,making the network achieve a strong balance between reducing the complexity of architecture and computation and improving the performance of the model.In addition,Distance-IoU loss is used to solve the problem of inaccurate regression position of bounding box and effectively speeds up network convergence.Finally,the number of parameters of the model was compressed to 1.54 MB less than YOLO Nano(4.0MB),and the mAP on the VOC2007 data set was 72.1%higher than the existing YOLO Nano(69.1%).

关 键 词：目标检测 轻量级 YOLO深度卷积神经网络 Pascal VOC 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]