【作者】 张强；

【导师】 张玉明； 扈啸；

【作者基本信息】 西安电子科技大学 ， 工程硕士（专业学位）， 2020， 硕士

【摘要】 伴随着近几年来人工智能的迅速发展,深度学习这一目前人工智能理论中最重要的技术被广泛的应用于各个领域之中,并起到了非常重要的作用。其中,计算机视觉一直是深度学习研究中最为活跃的几个研究方向之一,目标检测则是计算机视觉研究中的重点和难点。在军事领域中,目标检测在战场侦察、自动巡逻打击、无人作战和防御警戒等方面都有着重要的作用。由于深度学习涉及到很多向量和多矩阵运算,所以目前大量关于深度学习的研究都是基于GPU开展的,而GPU作为一个通用计算平台在应用部署中又存在功耗大的问题。本文主要研究了一种脱离GPU的深度学习算法应用,根据国防科技大学计算机学院自主研发的FT系列DSP的特点,结合对现有深度卷积神经网络的分析,开发了一个针对移动军事目标的深度学习方法目标检测系统,并且在FT-M7002多核DSP嵌入式平台上完成了部署及优化,实现了一个低功耗高性能的嵌入式深度学习应用。本文的主要工作如下:(1)建立了一个针对军事目标检测的小样本数据集。深度学习算法前期需要大量的数据进行反复迭代训练,而由于军事目标的特殊性,目前并没有公开的数据集可以直接使用。本文收集了6类公开的军事目标图像共6000张,并对其进行了数据标注。该数据集可以用来进行深度学习和计算机视觉相关研究训练的基础数据集。(2)在现有目标检测算法的基础上进行改进提升。本文对现在的深度学习目标检测算法进行了分析与总结,融合了先进的YOLO和Faster R-CNN思想,实现了一种更适用于FT系列DSP部署的目标检测网络。最终在训练集上达到了75%的正确识别率,高于YOLO的67%和Faster R-CNN的72%,在场景实测中普遍可以达到80%以上的识别率。(3)实现了一个小样本的深度学习训练。本文结合迁移学习等技术,实现了基于小样本的深度学习网络并成功部署在FT-M7002 DSP IMG图像系统中,为小样本深度学习应用研究提供了参考方法。(4)在FT系列嵌入式平台上优化了深度学习目标检测系统。现在深度学习应用落地的一大难点就在于其计算量的庞大导致需求算力较高,而相对应的通用计算平台的功耗和成本都难以控制。本文结合FT系列DSP的体系结构特点,对深度卷积神经网络的运算过程进行深度优化,在FT-M7002多核DSP嵌入式平台上完成了军事目标检测任务,充分的发挥了该嵌入式平台的低功耗和便于嵌入其他装备等优点,为深度学习算法在FT系列DSP上的应用提供了参考方法,并具有较高的实用价值。相比于先进的桌面级CPU intel i7-4900和深度学习训练常用的Nvdia GPU GTX 1080 Ti,在相同网络同等运算量的情况下,本文设计的DSP系统的功耗分别只有其17.8%和6.1%。更多还原

【Abstract】 With the rapid development of artificial intelligence in recent years,deep learning,the most important technology in current artificial intelligence theory,has been widely used in various fields and has played a very important role.Among them,computer vision has always been one of the most active research directions in deep learning research,and target detection is the focus and difficulty of computer vision research.In the military field,target detection plays an important role in battlefield reconnaissance,automatic patrol strikes,unmanned operations,and defensive vigilance.Because deep learning involves a lot of vector and multi-matrix operations,a large amount of research on deep learning is currently based on GPUs.As a general-purpose computing platform,there is a problem of high power consumption in application deployment.This paper mainly studies the application of a deep learning algorithm without GPU.According to the characteristics of the FT series DSP independently developed by the School of Computer Science and Technology of National University of Defense Technology,combined with the analysis of existing deep convolutional neural networks,a deep target for mobile military targets is developed.Learning method target detection system,and the deployment and optimization have been completed on the FT-M7002 multi-core DSP embedded platform,realizing a low-power and high-performance embedded deep learning application.The main work of this article is as follows:(1)A small sample data set for military target detection is established.In the early stage of deep learning algorithms,a large amount of data is required for repeated iterative training.Due to the special nature of military targets,there are currently no publicly available data sets that can be used directly.A total of 3,600 images of military targets in six categories are collected in this paper,and the data are labeled.This data set can be used as a basic data set for deep learning and computer vision related research training.(2)Improve and improve on the basis of existing target detection algorithms.This article analyzes and summarizes the current deep learning target detection algorithms,and combines advanced YOLO and Faster R-CNN ideas to achieve a target detection network that is more suitable for FT series DSP deployment.In the end,it achieved a correct recognition rate of 75% on the training set,which is higher than 67% of YOLO and 72% of Faster R-CNN.Generally,it can reach a recognition rate of more than 80% in scene measurement.(3)A small sample of deep learning training is implemented.This paper combines transfer learning and other technologies to implement a small sample-based deep learning network and successfully deploy it in the FT-M7002 DSP IMG image system,providing a reference method for small sample deep learning application research.(4)The deep learning target detection system is optimized on the FT series embedded platform.One of the major difficulties in the implementation of deep learning applications is that the huge amount of calculations leads to high demand for computing power,and the power consumption and cost of corresponding general-purpose computing platforms are difficult to control.Based on the architecture characteristics of FT series DSP,this paper deeply optimizes the operation process of deep convolutional neural network,completes the military target detection task on the FT-M7002 multi-core DSP embedded platform,and makes full use of the embedded platform’s The advantages of power consumption and ease of embedding in other equipment provide a reference method for the application of deep learning algorithms on FT series DSPs,and have high practical value.Compared with the advanced desktop-level CPU intel i7-4900 and the Nvdia GPU GTX 1080 Ti commonly used in deep learning training,the power consumption of the DSP system designed in this paper is only 17.8% and 6.1%,respectively,under the same computing load on the same network..更多还原