High-Performance Light Field Reconstruction with Channel-wise and SAI-wise Attention

Abstract

Light field (LF) images provide rich information and are suitable for high-level computer vision applications. To acquire capabilities of modeling the correlated information of LF, most of the previous methods have to stack several convolutional layers to improve the feature representation and result in heavy computation and large model sizes. In this paper, we propose channel-wise and SAI-wise attention modules to enhance the feature representation at a low cost. The channel-wise attention module helps to focus on important channels while the SAI-wise attention module guides the network to pay more attention to informative SAIs. The experimental results demonstrate that the baseline network can achieve better performance with the aid of the attention modules.