Light Field Images Compression Based on Graph Convolution Networks

Abstract

The light field records all the information about the light in space. With this property, it has a very good application prospect in immersive media. Due to the large amount of light field data, light field compression technology has attracted much attention. The graph data structure of vertices and edges can well describe the relationship between the light field viewpoints, which is suitable for processing light field data. At the same time, the graph convolution network (GCN) combines graph and neural networks and has great potential for processing graph data. In this work, a light field compression scheme based on the graph convolutional network has been proposed. On the compression side, the scheme will select the anchor views and the entire light field as the input and label of the GCN, which generates a network model. This model describes the relationship between the light field viewpoints. On the decompression side, the scheme reconstructs the entire light field using the model and the extracted anchor views. Experimental results show that the proposed scheme has superior performance in reconstructed images. Compared with state-of-the-art, it can achieve a gain of up to 4 dB on PSNR. Besides, the scheme has been proved to have good generalization capability.