| 摘要: |
| 光场（LF）信息具有高维特性，重建任务中所需要的空间信息与角度信息在宏像素图中高度耦合.为了充分利用空间角度信息，提高超分辨率（SR）重建质量，提出一个改进的基于光场空间角度解耦机制的LF图像角度SR重建网络设计.考虑到图像中的不同特征对重建质量的影响，通过不同的通道分配机制改变各特征的影响程度，提高重建准确性，在堆叠特征提取层的同时，引入注意力机制，获取更加丰富的空间角度信息.在测试场景上的实验结果表明，所提出的重建网络在合成与真实场景里都有较好的重建效果.在两个合成场景数据集上峰值信噪比/结构相似性（PSNR/SSIM）参数分别为34.62/0.964与42.68/0.972，在真实场景上的PSNR/SSIM均值为39.67/0.990. |
| 关键词: 光场(LF)图像 超分辨率(SR) 解耦机制 注意力机制 |
| DOI：10.3969/J.ISSN.1000-5137.2023.02.004 |
| 分类号:TP919.81 |
| 基金项目:福建省自然科学基金（2019J01816）；莆田市科技局资助项目（2021G2001-8）；福建省高等学校新世纪优秀人才项目（2018JYTRC（PU））；莆田学院引进人才科研启动资助项目（2019003） |
|
| Research on super-resolution reconstruction of light field images based on spatial angle information decoupling |
|
ZHANG Fang1, ZHANG Qian1, LIAO Wan1, LIU Faguo1, WANG Bin1, YAN Tao2
|
|
1.College of Information, Mechanical and Electrical Engineering, Shanghai Normal University, Shanghai 201418, China;2.School of Mechanical, Electrical and Information Engineering, Putian University, Putian 351100, Fujian, China
|
| Abstract: |
| Light field (LF) information had high-dimensional characteristics. The spatial information required in the reconstruction task was highly coupled with the angular information in the macro pixel map. To fully utilize the spatial angular information to improve the quality of super-resolution (SR) reconstruction, an improved LF image angular SR reconstruction network based on the LF spatial angular decoupling mechanism was proposed. Considering the influence of different features in the image on the reconstruction quality, the degree of influence from different features is modified by different channel assignment mechanisms to improve the reconstruction accuracy, and the attention mechanism was introduced to obtain richer spatial angle information while stacking feature extraction layers. The results of the experiments on the test scenes showed that the reconstruction network proposed in this paper had a better reconstruction effect on both synthetic and real scenes. The peak signal to noise ratio/structural similarity (PSNR/SIM) values were 34.62/0.964 and 42.68/0.972 on the two synthetic scene data sets, and the average PSNR/SIM values on the real scene were 39.67/0.990. |
| Key words: light field(LF) image super-resolution(SR) decoupling mechanism attention mechanism |
