Implicit Neural Representation for Light Field Image

Light field has emerged as one of the most compelling data types for immersive media. However, the massive data volume inherent to their multi-dimensional nature poses significant challenges for storage and transmission. To address this, this paper proposes an implicit neural representation-based method for light field compression. This method integrates temporal token attention and local feature extraction mechanisms to enhance the model's overall feature representation capability. In order to further enhance the reconstruction quality of the network decoding light field, this paper expands the size of the convolution kernel in the later stage of the network to increase the receiving field, to obtain as much feature information as possible. The proposed scheme was evaluated for compression performance on the EPFL light field dataset. Experimental results demonstrate that the proposed method could achieve comparative performance in terms of both rate-distortion performance and visual quality. Compared with the existing NeRV-based video compression baselines, the proposed method could achieve about 36.77% BD-rate savings over NeRV.