Light Field Compression via a Variational Graph Auto-Encoder

Massive light field (LF) data bring tremendous storage and transmission challenges, making the LF compression scheme highly demanded. This paper proposes a novel LF compression method via a variational graph auto-encoder (VGAE), aiming to exploit better the structural information of edges and vertices of the graph LF image. More specifically, the graph adjacency ma-trix and feature matrix are derived from the original graph data in the encoder. Subsequently, a graph convolutional network (GCN) is utilized to determine a multi-dimensional Gaussian distribution, from which the latent representation can be derived by sampling. Finally, the graph LF image can be reconstructed by the inner product of the latent variable in the decoder. The distinct charac-teristics of the proposed scheme lie in that VGAE encoder applies GCN as a function, which can better alleviate the loss of compression. Moreover, the divergence between the original and the reconstructed signals is evaluated using KL divergence to ensure that the estimator is unbiased, leading to better adaptability. The ex-perimental results demonstrate that the proposed method achieves better performance than the state-of-the-art methods.