兰州大学机构库 >数学与统计学院
应用于嵌入式平台的深度学习目标检测算法研究
Alternative TitleResearch on Deep Learning Object Detection Technique and its Applications in Embedded System
王兆男
Thesis Advisor赵学靖
2018-03-01
Degree Grantor兰州大学
Place of Conferral兰州
Degree Name硕士
Keyword深度学习 目标检测 焦点损失 模型压缩 嵌入式系统
Abstract

本文将以基于NVIDIA Jetson TX1的注视检测和基于iOS平台的商品检测为例讨论两种最先进的单步检测算法YOLO和SSD在嵌入式设备上进行实时目标检测的实现方法。其中涉及到深度神经网络的结构选择、优化策略、训练和模型压缩等方面的内容。对于YOLO算法,我们使用小网络结构,并通过引入焦点损失、使用卷积层剪枝算法等方式提高精度、压缩模型存储体积、提高模型运行速度。对于SSD算法,我们通过将传统卷积结构替代为 MobileNets 网络块结构的方式在保证网络深度的同时从根本上减少了模型参数,达到了高效低损模型压缩的目的。实验表明,焦点损失能够有效缓解目标检测任务中前景类、背景类损失不均衡的问题,有提升模型精度、提高收敛稳定性的作用;剪枝算法能够在大幅缩减模型尺寸的同时提高模型精度,代价是训练和调参过程较为繁琐;MobileNets网络结构能够在减小模型尺寸并将精度损失控制在可接受的范围内,而且训练方便快捷。

Other Abstract

This article will focus on two examples which are the Jetson TX1-based vision detection and iOS-based commodity detection to discuss two of the most advanced single-step detection algorithms YOLO and SSD in embedded devices for real-time object detection method.It would involve the structure selection of the deep neural network, optimization strategies, training and model compression and other aspects. For the YOLO algorithm, we use a small network as base structure, and improve the accuracy and reduce model volume by using Focal loss and pruning algorithm. For the SSD algorithm, we replace the traditional convolutional structure with the MobileNets structure to ensure the network depth while fundamentally reducing the model parameters, achieving efficient and low-loss model compression. Experiments shows that the Focal loss can effectively alleviate the problem of unbalanced foreground and background loss in the object detection task, improve model accuracy and improve convergence stability. The pruning algorithm can greatly reduce model size and improve model accuracy. The cost is that the training is more complex; the MobileNets can reduce the model size and control the loss of precision within an acceptable range, and the training is quick and easy.

URL查看原文
Language中文
Document Type学位论文
Identifierhttps://ir.lzu.edu.cn/handle/262010/224391
Collection数学与统计学院
Recommended Citation
GB/T 7714
王兆男. 应用于嵌入式平台的深度学习目标检测算法研究[D]. 兰州. 兰州大学,2018.
Files in This Item:
There are no files associated with this item.
Related Services
Recommend this item
Bookmark
Usage statistics
Export to Endnote
Altmetrics Score
Google Scholar
Similar articles in Google Scholar
[王兆男]'s Articles
Baidu academic
Similar articles in Baidu academic
[王兆男]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[王兆男]'s Articles
Terms of Use
No data!
Social Bookmark/Share
No comment.
Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.