Message Passing Neural Network based Light Field Image Compression

Light Field possesses the ability to capture comprehensive information about light distribution in space, making it an appealing technology for immersive media applications. However, due to the vast amount of data generated by the Light Field, there has been significant interest in developing methods for its compression. The use of a graph data structure that outlines the relationship between viewpoints within the Light Field has been found to be a suitable way to process this data. Additionally, the graph convolution network (GCN), which merges graph and neural network approaches, has shown great promise in effectively handling graph data. However, since GCN doesn't directly support edge features, it is unable to weigh the contribution of source key view features, aggregating at the target non-key view, in the graph representation of a Light Field data. In this work, a Light Field compression scheme based on Message Passing Neural Network (MPNN) is proposed to account for the edge features during the training and reconstruction of non-key views, which is found to perform well on a sparse down-sampling pattern of key views. Experimental results show a gain of about 1 dB in PSNR, compared with the previous GCN-based approach, and our proposal also outperformed other baselines at low bitrates, using an adaptive quantization parameters (QP) selection.