一种自适应权重学习的轻量超分辨率重建网络  

Lightweight image super-resolution with the adaptive weight learning network

作　　者：张宇浩 程培涛[1] 张书豪 王秀美[2] ZHANG Yuhao;CHENG Peitao;ZHANG Shuhao;WANG Xiumei(School of Electro-Mechanical Engineering,Xidian University,Xi’an 710071,China;School of Electronic Engineering,Xidian University,Xi’an 710071,China)

机构地区：[1]西安电子科技大学机电工程学院,陕西西安710071 [2]西安电子科技大学电子工程学院,陕西西安710071

出　　处：《西安电子科技大学学报》2021年第5期15-22,共8页Journal of Xidian University

基　　金：国家自然科学基金(61871308,61972305);陕西省重点研发计划(2021ZDLGY02-03)。

摘　　要：近年来,基于深度卷积神经网络的单幅图像超分辨率重建方法取得了令人瞩目的成果。基于像素级注意力网络的图像超分辨率重建方法能够在极小的参数量下获得良好的重建性能,是目前最先进的轻量化超分辨率重建方法之一。但是,受到各模块参数量的限制,像素级注意力网络训练缓慢和收敛条件苛刻的问题变得日益突出。针对这些问题,提出了一种基于自适应权重学习的轻量化超分辨率重建网络。该网络使用多个自适应权重模块组成非线性映射网络,每个模块能够提取到不同层级的特征信息,在每个自适应权重模块中,利用注意力分支和无注意力分支分别获取相应信息,再通过自适应权重融合分支进行整合。使用特定的卷积层拆分和融合两条分支,大幅降低了注意力分支和无注意力分支的参数量,使网络在参数量与性能之间达到相对平衡。在标准数据集上的实验证明,所提出方法在降低模型参数量的同时,峰值信噪比和结构相似度两种客观质量评价指标均优于同类先进方法,该方法能够重建更准确的纹理细节,得到更好的视觉效果,证明了该方法的有效性。In recent years,the single-image super-resolution(SISR)method using deep convolutional neural networks(CNN)has achieved remarkable results.The Pixel Attention Network(PAN)is one of the most advanced lightweight super-resolution methods,which can lead to a good reconstruction performance with a very small number of parameters.But the PAN is limited by the parameters of each module,resulting in slow model training and strict training conditions.To address these problems,this paper proposes a Lightweight Adaptive Weight learning Network(LAWN)for image super-resolution.The network uses multiple adaptive weight modules to form a non-linear mapping network,with each module extracting different levels of feature information.In each adaptive weight module,the network employs the attention branch and the non-attention branch to extract the corresponding information,and then the adaptive weight fusion branch is employed to integrate these two branches.Splitting and fusing the two branches with a specific convolutional layer greatly reduces the number of parameters of the attention branch and the non-attention branch,which helps the network to achieve a relative balance between the number of parameters and the performance.The quantitative evaluations on benchmark datasets demonstrate that the proposed LAWN reduces the number of model parameters and performs favorably against state-of-the-art methods in terms of both PSNR and SSIM.Experimental results show that this method can reconstruct more accurate texture details.The qualitative evaluations with better visual effects prove the effectiveness of the proposed method.

关 键 词：超分辨率重建 卷积神经网络 深度学习 轻量化 自适应权重 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]