面向边缘端的轻量级实时目标检测网络及VLSI加速方法研究
结题报告
批准号:
62002134
项目类别:
青年科学基金项目
资助金额:
24.0 万元
负责人:
骆爱文
依托单位:
学科分类:
计算机系统结构与硬件技术
结题年份:
2023
批准年份:
2020
项目状态:
已结题
项目参与者:
骆爱文
国基评审专家1V1指导 中标率高出同行96.8%
结合最新热点,提供专业选题建议
深度指导申报书撰写,确保创新可行
指导项目中标800+,快速提高中标率
客服二维码
微信扫码咨询
中文摘要
以移动终端为代表的边缘设备可以给目标检测系统带来无与伦比的可访问性,具有极高的实用价值。但现有的目标检测网络对硬件资源依赖严重,计算复杂度高、运行时间长、功耗高,难以部署在资源受限的边缘端的设备上,并满足其时效和能耗要求。目标检测旨在检测图像的前景目标对象框位置与识别背景所对应的标签类别。基于深度学习的目标检测方法因具有极高的检测精度而被广泛研究。本项目针对边缘端的实时应用场景,拟提出一种轻量级实时目标检测网络,降低网络参数规模并保证目标检测精度;通过设计合理的数据流及其硬件架构,在卷积计算、数据复用与存储等方面进行布局优化和硬件电路集成化设计,以提高目标检测速度,降低硬件资源消耗和功率损耗。本项目采用硬件电路加速的新思路,预期研究成果可为实时目标检测系统的基础研究和应用研究作出积极贡献。
英文摘要
The edge device represented by the mobile terminal can bring unprecedented accessibility to the object detection system, which brings extremely high value to the application scenarios. However, the existing neural networks rely heavily on computing hardware resources, leading to the results of high computational complexity, long running time and high power consumption. It is quite difficult to deploy the current neural networks on the time-sensitive edge terminals with limited resources and meet their energy requirements. Object detection is to find out the foreground target object box position of the image and recognize the label category corresponding to the background. The deep-learning-based detection approaches have been widely studied because of its high accuracy. Focusing on the real-time application scenario of edge terminals, this research intends to develop a lightweight convolutional neural network for real-time object detection that is suitable for deployment on edge devices. The network parameters will be reduced with the guarantee of holding a comparable accuracy. The computing procedure will be accelerated by constructing an optimized hardware architecture for the data flow. The acceleration structure is optimized in the aspects of the convolutional calculation, data reuse, and storage scheme, etc., to reduce hardware resource requirements and power consumption. This research employs a new idea of using hardware-oriented acceleration for object detection, whose results will make a great contribution to the fundamental research and applied research of the real-time image detection system on neural networks.
期刊论文列表
专著列表
科研奖励列表
会议论文列表
专利列表
DOI:10.1109/jsen.2021.3059099
发表时间:2021-04
期刊:IEEE Sensors Journal
影响因子:4.3
作者:A. Luo;S. Bhattacharya;S. Dutta;Y. Ochi;M. Miura-Mattausch;Jian Weng;Yicong Zhou;H. Mattausch-H.-Mattau
通讯作者:A. Luo;S. Bhattacharya;S. Dutta;Y. Ochi;M. Miura-Mattausch;Jian Weng;Yicong Zhou;H. Mattausch-H.-Mattau
DOI:10.1109/les.2023.3299114
发表时间:2024
期刊:IEEE Embedded Systems Letters
影响因子:1.6
作者:Aiwen Luo;S. Bhattacharya;M. Miura;Yicong Zhou;H. Mattausch
通讯作者:H. Mattausch
DOI:10.1109/tnnls.2022.3176493
发表时间:2022-05
期刊:IEEE Transactions on Neural Networks and Learning Systems
影响因子:10.4
作者:Min Shi;Jialin Shen;Qingming Yi;Jian Weng;Zunkai Huang;Aiwen Luo;Yicong Zhou
通讯作者:Min Shi;Jialin Shen;Qingming Yi;Jian Weng;Zunkai Huang;Aiwen Luo;Yicong Zhou
DOI:10.1007/s11063-023-11145-z
发表时间:2023-01
期刊:Neural Processing Letters
影响因子:3.1
作者:Qingming Yi;Guoshuai Dai;Min Shi;Zunkai Huang;Aiwen Luo
通讯作者:Qingming Yi;Guoshuai Dai;Min Shi;Zunkai Huang;Aiwen Luo
DOI:10.3778/j.issn.1673-9418.2203015
发表时间:2022
期刊:计算机科学与探索
影响因子:--
作者:石敏;沈佳林;易清明;骆爱文
通讯作者:骆爱文
国内基金
海外基金