Light Field Image Coding using High Order Prediction Training

This paper proposes a new method for light field image coding relying on a high order prediction mode based on a training algorithm. The proposed approach is applied as an Intra prediction method based on a two-stage block-wise high order prediction model that supports geometric transformations up to eight degrees of freedom. Light field images comprise an array of micro-images that are related by complex perspective deformations that cannot be efficiently compensated by state-of-the-art image coding techniques, which are usually based on low order translational prediction models. The proposed prediction mode is able to exploit the non-local spatial redundancy introduced by light field image structure and a training algorithm is applied on different micro-images that are available in the reference region aiming at reducing the amount of signaling data sent to the receiver. The training direction that generates the most efficient geometric transformation for the current block is determined in the encoder side and signaled to the decoder using an index. The decoder is therefore able to repeat the high order prediction training to generate the desired geometric transformation. Experimental results show bitrate savings up to 12.57% and 50.03 % relatively to a light field image coding solution based on low order prediction without training and HEVC, respectively.

I. Introduction

Standard cameras are composed of two main elements: the lens and the camera sensor, which allows to capture light hitting the camera sensor on specific spatial coordinates. In single tier lenslet Light Field (LF) cameras, a third element is added: the microlens array (MLA). The MLA allows the LF camera to also capture the angular information of light hitting the sensor [1]. Each microlens creates a micro-image (MI) on the sensor containing both spatial and angular information about the light hitting the sensor. Thus, the captured LF image contains 3D information about the scene, instead of a single 2D perspective.