一种基于对极约束注意力的光场图像超分辨率方法  

An epipolar constraint attention method for light field super resolution

作　　者：杨敬钰 金满昌 胡昆澍 刘高生 YANG Jingyu;JIN Manchang;HU Kunshu;LIU Gaosheng(School of Electrical and Information Engineering,Tianjin University,Tianjin 300072,China)

机构地区：[1]天津大学电气自动化与信息工程学院,天津300072

出　　处：《华中科技大学学报(自然科学版)》2025年第3期1-7,共7页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：国家自然科学基金资助项目(62231018)。

摘　　要：针对光场成像中因空间-角度复用,导致空间分辨率成角度倍下降的问题,提出了一种基于对极约束注意力机制的深度学习光场空间超分辨率方法.通过分析光场的视差分布特性,结合深度学习注意力网络的原理,设计了针对光场图像处理的对极约束注意力模块,将注意力限制在了光场对极平面图像的一个特定范围内,从而强化光场图像超分辨率的相关特征和抑制非相关特征.实验结果表明:该范围在和输入光场的视差范围匹配时,可以达到最好的性能,具备一定的可解释性,所提出的方法在平均峰值信噪比和结构相似性等指标上优于10种比较方法,并且具有更好的视觉效果.To solve the spatial resolution limitations of light field imaging,caused by spatial-angular multiplexing and lead to spatial resolution degradation,a deep learning-based light field spatial super-resolution method was proposed,utilizing an epipolar constraint attention mechanism.By analyzing the disparity distribution characteristics of the light field and the combination of the principle of attention networks,an epipolar constraint attention module was designed to confine attention within a specific range of the light field's epipolar-plane image,thereby enhancing light field super-resolution relevant features and suppressing irrelevant ones.Experimental results demonstrate that this range can achieve the best performance when matched with the disparity range of the input light field,with a certain degree of interpretability.The experimental results also show that the proposed method outperforms 10comparative methods in terms of average peak signal-to-noise ratio and structural similarity and exhibits superior visual quality.

关 键 词：光场图像 图像超分辨率 对极约束 深度学习 注意力机制 

分 类 号：TN911.73[电子电信—通信与信息系统]