摘要
近年来基于深度卷积神经网络的单幅图像超分辨率技术取得了很大进展.但特征提取方式单一,模型参数量大很难在移动端部署.为了解决这些问题,本文提出了一种多分支残差特征蒸馏算法.首先,通过多分支残差模块进行深层特征提取;其次,结合卷积、通道自适应激活函数和瓶颈注意力模块进行特征蒸馏及融合,减少平坦区域的大量冗余参数,在保证性能的同时降低模型复杂度;最后通过亚像素卷积层进行图像重建,得到最终的超分辨率图像.实验结果表明该算法在模型复杂度和性能上达到更好的平衡.与IMDN(Information Multi-distillation Network)相比,该算法的PSNR和SSIM分别有0.06~0.26dB与0.001~0.006的提升;在2倍超分重建结果中,与千万级参数量模型DBPN(Deep Back-Projection Networks)相比,本文算法参数量是其1/15,PSNR基本相同,SSIM提高0.001.
Recently,deep convolutional neural networks has made great progress in single image super-resolution.But the feature extraction method is single and it is not easy to apply these methods to edge devices due to the requirement of heavy computation.In order to solve these problems,this paper proposes a diverse branch residual feature distillation algorithm.Firstly,deep feature extraction is performed through the diverse branch residual block.Then,feature distillation and fusion are performed through convolution,channel adaptive activation function,and bottleneck attention module to reduce a large number of redundant parameters in the flat area,so it can reduce model complexity while ensuring performance.Finally,the image is reconstructed by the sub-pixel convolutional layer to obtain the final super-resolution image.Experimental results show that the algorithm achieves a better balance between model complexity and performance.Compared with IMDN(Information Multi-distillation Network),the PSNR and SSIM of this algorithm are improved by 0.06-0.26dB and 0.001-0.006 respectively.In the 2x super-resolution reconstruction results,compared with the tens of millions of parameter model DBPN(Deep Back-Projection Networks),the algorithm parameter in this paper is 1/15 of it,PSNR is basically the same,and SSIM is improved by 0.001.
作者
李轩
刘立柱
LI Xuan;LIU Li-zhu(College of Electronic and Information Engineering,Shenyang Aerospace University,Shenyang 110136,China)
出处
《小型微型计算机系统》
CSCD
北大核心
2023年第2期363-369,共7页
Journal of Chinese Computer Systems
基金
国家自然科学基金项目(61901284)资助
辽宁省“兴辽英才计划”项目(XLYC1907022)资助
辽宁省重点研发计划项目(2020JH2/10100045)资助
关键词
图像超分辨率
多分支卷积
残差模块
注意力机制
特征蒸馏
image super-resolution
diverse branch convolution
residual module
attention mechanism
feature distillation
