摘要
针对工业应用领域中终端设备计算能力较低且对检测算法的响应速度存在较高需求的问题,提出基于分裂倒残差的轻量型实时目标检测算法.首先,在主干网络中使用分裂倒残差结构,削减网络结构的参数量以及运算次数,以达到加快推理速度的目的;其次,引入自适应上下文感知模块以及轻量型双向特征融合模块,旨在提升特征信息交流、增加对小目标检测性能的同时,避免增加额外的学习参数与推理.实验结果表明,文中算法在参数量仅有7.5×105的情况下,MS COCO数据集中检测精度达到21.1%,移动端检测速度达到48帧/s,远超对比算法,该检测算法更适合在无法提供高计算能力的移动端设备上完成目标检测任务.
For the problem of low computing power of terminal equipment and high demand for the response speed of detection algorithms in the industrial application field,a lightweight real-time object detection algorithm based on split inverted residuals is proposed.Firstly,the split inverted residuals are used in the backbone network to reduce the number of parameters and calculations to achieve faster inference.Secondly,the self-adaptive context-awareness module and the lightweight two-way feature fusion module are introduced to improve the feature information exchange and increase the detection performance of small targets while avoiding the addition of learning parameters and inference cost.According to the results of experiments,the algorithm has a detection accuracy of 21.1%in the MS COCO data set and a detection speed of 48 frame/s,far exceeding the comparison when the parameter amount is only 0.75M.The detection algorithm is more suitable for object detection tasks on mobile devices which can not provide high computing power.
作者
周浩然
侯进
杨宗源
曾雷鸣
康萍萍
Zhou Haoran;Hou Jin;Yang Zongyuan;Zeng Leiming;Kang Pingping(IPSOM Lab,School of Information Science and Technology,Southwest Jiaotong University,Chengdu 611756;School of Computer and Artificial Intelligence,Southwest Jiaotong University,Chengdu 611756;National Engineering Laboratory of Integrated Transportation Big Data Application Technology,Chengdu 611756;Tangshan Institute,Southwest Jiaotong University,Tangshan 063000)
出处
《计算机辅助设计与图形学学报》
EI
CSCD
北大核心
2023年第1期66-74,共9页
Journal of Computer-Aided Design & Computer Graphics
基金
国家重点研发计划(2020YFB1711902)。
关键词
深度学习
目标检测
残差结构
双向特征融合
deep learning
object detection
residual structure
two-way feature fusion
