基于组卷积特征融合的One-Stage目标检测模型  

One-Stage Target Detection Model Based on Group Convolution Feature Fusion

作　　者：鲍先富 强赞霞[1] 李丹阳 杨瑞 BAO Xian-fu;QIANG Zan-xia;LI Dan-yang;YANG Rui(Zhongyuan University of Technology,Zhengzhou 450007,China)

机构地区：[1]中原工学院,河南郑州450007

出　　处：《计算机技术与发展》2021年第11期86-94,共9页Computer Technology and Development

基　　金：河南省科技计划项目(182102210126)。

摘　　要：由于移动终端计算能力和内存大小的限制,在移动设备上进行实时目标检测具有非常大的挑战性。为了更好地在无人驾驶汽车等移动设备上进行目标检测,该文以YOLOv3单阶段目标检测模型为基础,对部署在移动设备上的目标检测模型进行优化,以提高模型的检测精度(MAP)并降低计算复杂度。具体改进措施为:基于DarkNet-53为主干网络引入组卷积和通道洗牌技术;基于M.G.Hluchyj等学者提出的网络设计指导原则,对主干网络的残差单元和下采样单元进行修改优化;为减轻YOLOv3模型对于密集目标的漏选和标签重写问题,引入特征混合金字塔模型。通过在Pascal VOC2007和VOC2012数据集上进行实验对比,优化模型的整体精度较YOLOv3提高8.17%,模型参数量降低1.21 M,在与YOLOv4的参数量大体相等的情况下达到了YOLOv4的检测精度。Due to the limitations of mobile terminal computing power and memory size,real-time target detection on mobile devices is quite challenging.In order to better perform target detection on mobile devices such as driverless cars,based on YOLOv3 single-stage target detection model,the target detection model deployed on the mobile device is optimized to improve the detection accuracy(MAP)and reduce the computational complexity.The specific improvement measures are as follows:the introduction of volume and channel shuffling technology based on the DarkNet-53 backbone network;based on the network design guidelines proposed by scholars such as MG Hluchyj,the residual unit and down-sampling unit of the backbone network are modified and optimized;the feature mixture pyramid model is introduced to reduce the YOLOv3 model’s omission of dense targets and label rewriting.Through experimental comparison on the Pascal VOC2007 and VOC2012 data sets,the overall accuracy of the optimized model is 8.17%higher than that of YOLOv3,and the model parameter is reduced by 1.21 M.The detection accuracy of YOLOv4 is reached when the parameter quantity of YOLOv4 is roughly equal.

关 键 词：卷积神经网络 目标检测 残差网络 特征融合金字塔 通道洗牌 组卷积 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]