【作者】 李琦；

【导师】 郭炜； 周津；

【作者基本信息】 天津大学 ， 计算机技术， 2019， 硕士

【摘要】 由于现在的图像质量越来越高,现有的图像压缩技术已经不能满足目前的需求。于是使用卷积神经网络解决图像压缩成为目前热门的研究领域。但是,通过卷积神经网络训练学习后得到的网络模型不管在存储还是计算上都耗费了庞大的计算机资源。如何解决这个问题成为工业界和学术界研究的重点。本文提出一种深度卷积生成式对抗网络结构,该网络模型由卷积网络构成,可以带来比现有技术更加好的压缩效果,但是复杂的网络模型与巨大的计算量带来了庞大的计算机资源开销。于是,本文提出了一种网络模型的优化方法,在保证不损失网络模型原始精度且降低资源开销的前提下,通过修剪每层网络中被认为对输出精度影响很小的卷积核的方法来降低资源开销。实验表明,可以在确保原始精度的同时将原网络模型的存储降低约75%,计算量降低约50%。本文又尝试将移动网络的结构特点融合到该模型中,实验结果显示可以将该网络模型的存储降低约85%。更多还原

【Abstract】 Due to the increasing image quality,the existing image compression technology can not meet the current needs.Therefore,it has become a hot research field that the use of convolutional neural networks to solve image compression.However,the network model obtained through the training of convolutional neural networks consumes huge computer resources both in storage and in computation.How to solve this problem has become the focus of research in industry and academia.This thesis proposes a deep convolution generator adversarial networks structure.The network model consists of convolution neural networks,which can bring better compression effects than the prior art,but the complex network model and amount of computing bring huge computer resource overhead.Thus,this thesis proposes a method for optimizing the network model.Under the premise of ensuring that the original accuracy of the network model is not lost and the resource overhead is reduced,the resource overhead is reduced by trimming the convolution kernel in each layer network which is considered to have little influence on the output precision.Experiments have shown that the original network model can be reduced by about 75% and the computational amount by about 50% while ensuring the original accuracy.In this thesis,the structural characteristics of the mobile network are merged into the model.The experimental results show that the storage of the network model can be reduced by about 85%.更多还原