多维注意力机制与选择性特征融合的图像超分辨率重建  

Multidimensional attention mechanism and selective feature fusion for image super-resolution reconstruction

作　　者：温剑 邵剑飞[1] 刘杰 邵建龙[1] 冯宇航 叶榕 WEN Jian;SHAO Jianfei;LIU Jie;SHAO Jianlong;FENG Yuhang;YE Rong(Faculty of Information Engineering and Automation,Kunming University of Science and Technology,Kunming 650500,China;Yunnan Police Unmanned System Innovation Research Institute,Yunnan Police Officer Academy,Kunming 650223,China)

机构地区：[1]昆明理工大学信息工程与自动化学院,云南昆明650500 [2]云南警官学院云南警用无人系统创新研究院,云南昆明650223

出　　处：《光学精密工程》2023年第17期2584-2597,共14页Optics and Precision Engineering

基　　金：国家重点研发计划资助项目(No.2022YFC3320800);国家自然科学基金资助项目(No.61732005,No.61302042);云南省重大科技专项计划资助项目(No.202002AD080001)。

摘　　要：针对图像超分辨率重建过程中提取低分辨率特征效果较差,大量高频信息丢失导致的边缘模糊和伪影问题,提出了融合多维注意力机制与选择性特征融合作为图像特征提取模块的图像超分辨率重建方法。网络由若干个基本块和残差操作构建模型的特征提取结构,其核心是一种提取图像特征的异构组卷积块,该模块的对称组卷积块以并行的方式进行卷积提取不同通道间的内部信息特征并进行选择性特征融合,互补卷积块通过全维度动态卷积从空域、输入输出维度和核维度捕捉遗漏的上下文信息,对称组卷积块和互补卷积块连接后的特征采用特征增强残差块去除冗余造成干扰的无用信息。模型通过5种消融实验证明其设计的合理性,在Set5,Set14,BSDS100和Urban100测试集上与其他主流的超分辨率重建方法进行对比,峰值信噪比(PSNR)和结构相似性(SSIM)定量数据均有提升,尤其在放大因子为3的Set5数据集上比次优算法CARN-M均提升0.06 dB,结果表明提出模型具有更优的性能指标和更好的视觉效果。To address the problems of poor extraction of low-resolution features and blurred edges and arti⁃facts caused by the high loss of high-frequency information in an image super-resolution reconstruction pro⁃cess,this paper proposes an image super-resolution reconstruction method that combines multidimensional attention and selective feature fusion(SKFF)as an image feature extraction module.The network com⁃prises several basic blocks and residual operations to construct the feature extraction structure of the mod⁃el,the core of which is a heterogeneous group convolution block for extracting image features.The sym⁃metric group convolution block of this module performs convolution in a parallel manner to extract the in⁃ternal information between different feature channels and performs selective feature fusion.The comple⁃mentary convolution block captures the missed contextual information from the null domain,input–output dimension,and kernel dimension by full-dimensional dynamic convolution(ODconv).The features ob⁃tained after the symmetric group convolution and complementary convolution block processes are connect⁃ed via a feature-enhanced residual block to remove useless information causing interference by redundancy.The rationality of the model design is demonstrated through five ablation experiments.Peak signal-tonoise ratio(PSNR)and structural similarity(SSIM)quantitative data comparison with other mainstream super-resolution reconstruction methods on the Set5,Set14,BSDS100,and Urban100 test sets are im⁃proved,especially on the Set5 dataset with an amplification factor of 3,showing a 0.06 dB improvement over the CARN-M algorithm.The experimental results demonstrate that the proposed model has better performance indexes and visual effects.

关 键 词：超分辨率重建 多维注意力机制 特征融合 残差网络 

分 类 号：TP394.1[自动化与计算机技术—计算机应用技术]