【作者】 王琪；

【导师】 吴海；

【作者基本信息】 华中科技大学 ， 计算机技术， 2019， 硕士

【摘要】 深度神经网络是近年来人工智能领域一大重要研究方向。目前,深度学习中的目标检测方法通常采用卷积神经网络模型,通过矩阵计算实现卷积神经网络的核心——卷积运算。卷积神经网络模型的推理需要较多的存储和计算资源。嵌入式设备中的存储和计算资源往往有限,传统的深度学习方法无法直接应用于此类系统环境。在推理过程中,由于网络的参数已经固定,一方面不再需要冗余的参数来保证网络的收敛,另一方面网络参数的冗余性还会大幅度降低网络前向传播的速度并占用额外的存储空间。因此,如何在训练完成后结合卷积神经网络的自身特点消除网络参数的冗余性并将该方法应用于目标检测任务上,是一个亟待解决的重要问题。为了解决这个问题,在前向传播过程中需要同时减少网络的参数数量和网络推理所需要的时间。主要工作如下:首先,对于卷积神经网络的基本原理与实现进行了简单分析,并介绍了基于该技术进行的目标检测任务的相关技术实现细节,为后续工作的展开提供了理论基础。其次,总结分析了目前主流的神经网络压缩和加速的方法,并对每一类方法的基本技术原理给出了简要的介绍,分析和比较了了相关方法的优缺点,为后续的工作指明了技术方向。最后,针对目前常用的基于卷积神经网络的目标检测网络,结合相关文献提出了针对卷积层进行稀疏化通道剪枝,对目标检测网络进行压缩和加速,并在嵌入式平台Nvidia jetson TX2上进行测试。实验过程使用了开源的深度学习框架平台pytorch实现了针对网络的压缩和加速,并使用Caffe平台进行了压缩和加速后的速度测试,验证了方法的有效性。更多还原

【Abstract】 Deep neural network is an important research field of artificial intelligence in recent years.Usually,object detection is based on convolutional neural network now.And the basis of convolution operation is matrix computation.Inference progress consumes large amounts of storage and computation resources.But they are limited in embedded devices.Traditional deep learning methods are hard to be used under these circumstances.During inference,parameter values are fixed.On one hand,we don’t need the redundancy,on the other hand,the redundancy consumes extra storage and slows the progress.So,it’s an important question to eliminate the redundancy of our networks and make it available for embedded object detection environment.To achieve this goal,shrinking parameters’ amount and accelerating inference are all needed.The main works are as below: Give a basic theory of convolutional network and its implementation,and some basic technology of convolutional network based object detection.A summary of mainstream neural network’s compression and acceleration method now is given.Give some basic idea of their mechanisms and implementations,and compare these method between their pros and cons.Propose a method for CNN-based object detection to compress and accelerate it by sparse channel pruning method from related works.And test this method on an embedded platform Nvidia jetson TX2 with an open source deep learning architecture pytorch.The speed before and after compression and acceleration on Caffe platform are tested.The effectiveness is proved by the project.更多还原