基于穿插特征更新的光场图像超分网络  

Cross Feature Updating-Based Network for Light Field Image Super Resolution

作　　者：陈佳妮 武迎春[1] 吕天琪 刘丽[1] 赵贤凌[1] CHEN Jia-ni;WU Ying-chun;LÜTian-qi;LIU Li;ZHAO Xian-ling(School of Electronic and Information Engineering,Taiyuan University of Science and Technology,Taiyuan,shanxi 030024,China)

机构地区：[1]太原科技大学电子信息工程学院,山西太原030024

出　　处：《电子学报》2024年第12期4113-4124,共12页Acta Electronica Sinica

基　　金：国家自然科学基金(No.61601318);山西省自然科学基金(No.202103021224278);山西省回国留学人员科研项目(No.2020-128)。

摘　　要：针对光场图像空间分辨率低的问题,本文搭建了基于穿插特征更新的光场图像超分网络,获得更高质量的光场子孔径图像阵列.网络的浅层特征提取部分采用3分支结构,设计了并行残差块从不同形式的光场数据中提取空间特征和角度特征.深层特征提取采用穿插特征更新结构,设计了特征对齐交互模块、自注意力特征交互模块和空间特征增强模块,实现了空间特征与角度特征逐级融合更新.数据重构部分通过交替使用多尺度残差块和通道注意力块将逐级更新后的空间信息融合,最终经数据上采样得到超分图像.所提网络在充分挖掘、补充空间、角度特征的基础上,采用逐级融合、更新、增强机制,实现不同层次空间信息的收集,获得更好的超分效果.对比实验验证了所提网络的优越性,在5组公开光场数据集上,本文搭建的网络在4倍超分任务下平均峰值信噪比(Peak Signal-to-Noise Ratio, PSNR)值达到32.31 dB,较现有网络表现出更好的超分性能.Addressing the issue of low spatial resolution in light field image,cross feature updating-based network for light field image super resolution is built in this paper to generate a higher quality array of light field sub-aperture imag-es.In this work,a 3-branch structure is adopted for shallow feature extraction.In order to extract spatial and angular fea-tures from different forms of light field data,parallel residual block is designed.A cross feature update structure is used for extract deep features,and feature alignment interaction module,self-attention feature interaction module,and spatial feature enhancement module are designed to achieve step-by-step fusion and updating of spatial and angular features.In the data re-construction part,the updated spatial information is fused by using multi-scale residual block and channel attention block al-ternately and finally super-resolution images are obtained through data upsampling.On the basis of fully exploring and sup-plementing spatial and angular features,the proposed network adopts a step-by-step fusion,update,and enhancement mech-anism to gather spatial information at different levels,leading to superior super-resolution results.Comparative experiments demonstrate the effectiveness of the proposed method,with the network achieving an average peak signal-to-noise ratio(PSNR)value of 32.31 dB for 4×tasks on 5 public light field datasets,surpassing the performance of existing networks.

关 键 词：4D光场 超分辨率重建 穿插特征更新 特征对齐交互 自注意力特征交互 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]