Light field (LF) imaging has proven to be a promising technique in computer vision field. However, there is a tradeoff between spatial and angular resolution of LF images, which limits the application of LF cameras. Super-resolution (SR) of the angular domain is proposed to improve the angular resolution of LF images. However, most of the SR frameworks cannot adapt to LF datasets with multi-size disparities, especially the large disparities. In this paper, we proposed a learning-based SR framework named EASRnet. The EASRnet consists of three parts - Disparity adaptation, Feature extraction, and Feature restoration parts, and achieves angular SR tasks by using residual blocks and a structure with branches to reconstruct high-frequency details of up-sampled epipolar plane images (EPI). It employs an additional blur layer to accommodate LF datasets with different disparities. The experimental results show that the proposed approach can reconstruct novel view images with satisfactory accuracy.
|