基于改进卷积神经网络的图像超分辨率算法研究  

Research on image super-resolution algorithm based on improved convolutional neural network

作　　者：胡晓辉[1] 张建国 Hu Xiaohui;Zhang Jianguo(School of Electronics&Information Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China)

机构地区：[1]兰州交通大学电子与信息工程学院,兰州730070

出　　处：《计算机应用研究》2020年第3期947-950,956,共5页Application Research of Computers

基　　金：国家自然科学基金资助项目(61163009);甘肃省科技计划资助项目(144NKCA040)。

摘　　要：针对现有卷积神经网络图像超分辨率复原算法中映射函数容易出现过学习、损失函数收敛性不足等问题,通过结合现有视觉识别算法和深度学习理论对其进行改进。首先将原有SRCNN层数从3层提高到13层,并提出一种自门控激活函数形式swish,代替以往网络模型常用的sigmoid、ReLU等激活函数,充分利用swish函数的优势,有效避免了过拟合问题,更好地学习利用低分辨率到高分辨率图像之间的映射关系指导图像重建;然后在传统网络损失函数中引入Newton-Raphson迭代法理论,进一步加快了收敛速度。最后通过实验证明了改进的卷积神经网络模型能够有效改善图像的清晰度,并在主观视觉效果和客观参数评价指标上有进一步提高。Aimed at the problems of over-fitting of mapping function and insufficient convergence of loss function in convolution neural network image super-resolution algorithm,combined existing vision recognition algorithm and depth learning theory,this paper proposed an improvement on it.Firstly,the original SRCNN increased layer number from 3 to 13 layers,and proposed a form of self-gated activation function swish to replace the usual network model sigmoid,ReLU and other activation functions,and fully utilized the advantages of swish function to effectively avoid over-fitting problems,better to learn and use the mapping relationship between low-resolution and high-resolution images to guide image reconstruction.Then it introduced the Newton-Raphson method into the traditional network loss function,which further accelerated the convergence speed.Finally,experiments show that the improved network model can effectively improves the image definition,and improve the visual effect and objective parameter evaluation index.

关 键 词：低分辨率 超分辨率 卷积神经网络 图像处理 复原 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]