Air-Net:一种轻量级的车辆检测方法  

Air-Net:A Lightweight Vehicle Detection Method

作　　者：胡序洋 高尚兵 李杰[1,2] 张秦涛 HU Xuyang;GAO Shangbing;LI Jie;ZHANG Qintao(School of Computer and Software Engineering,Huaiyin Institute of Technology,Huai’an 223001,China;Laboratory for Internet of Things and Mobile Internet Technology of Jiangsu Province,Huaiyin Institute of Technology,Huai’an 223001,China)

机构地区：[1]淮阴工学院计算机与软件工程学院,江苏淮安223001 [2]淮阴工学院、江苏省物联网移动互联网技术工程实验室,江苏淮安223001

出　　处：《江苏海洋大学学报（自然科学版）》2023年第1期44-50,共7页Journal of Jiangsu Ocean University:Natural Science Edition

基　　金：国家自然科学基金面上资助项目(62076107);国家重点研发计划项目(2018YFB1004904);江苏高校“青蓝工程”项目;江苏高校自然科学研究重大项目(18KJA520001)。

摘　　要：车辆检测是计算机视觉领域中最具挑战性的任务之一,提升算法检测的速度及准确率具有重要的现实意义。针对目前车辆检测算法参数量大、速度较慢问题,提出一种轻量级的车辆检测算法Air-Net。首先基于非对称卷积以及残差连接原理构建出多残差非对称卷积模块(MulRes-AC block)用于提取目标特征,非对称卷积能极大减少算法的参数量,多残差结构能缓解网络梯度消失与爆炸问题,大幅度提升网络性能。然后使用特征统一融合与分配模块(feature unified fusion and assignment module,FUFA)融合来自主干网络中不同尺度的特征信息,并使用通道注意力机制去除与检测任务无关的噪声信息,使检测器能够直接接收到多种尺度的特征信息。在车辆数据集上进行实验,该方法的mAP为92%,检测速度为49.8 f/s。实验结果表明,所提出的轻量级车辆检测算法能够满足边缘设备的检测精度以及实时性要求。Vehicle detection is one of the most challenging tasks in the field of computer vision.It is of great practical significance to improve the speed and accuracy of algorithm detection.In view of the large number of parameters and slow speed of the current vehicle detection algorithm,a lightweight vehicle detection algorithm air net is proposed.Firstly,based on the asymmetric convolution and residual connection principle,a MulRes-AC block is constructed to extract target features.Asymmetric convolution can greatly reduce the parameters of the algorithm.The multi residual structure can alleviate the problem of network gradient disappearance and explosion,and greatly improve the network performance.Then the feature unified fusion and assignment module(FUFA)is used to fuse the feature information of different scales from the backbone network,and the channel attention mechanism is used to remove the noise information irrelevant to the detection task,so that the detector can directly receive the feature information of multiple scales.The experimental results on vehicle data sets show that the mAP of this method is 92%and the detection speed is 49.8 f/s.Moreover,the proposed lightweight vehicle detection algorithm can meet the detection accuracy and real-time requirements of edge devices.

关 键 词：车辆检测 计算机视觉 注意力机制 特征融合 多残差结构 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]