Perceptual Light Field Image Coding with CTU Level Bit Allocation

Light field imaging simultaneously records the position and direction information of light in scene, as one of the important techniques for digital media. The amount of light field image (LFI) data is huge, it needs to be effectively compressed. In this paper, a perceptual LFI coding method with coding tree unit (CTU) level bit allocation strategy is proposed. To remove angular redundancy, a hybrid coding framework with joint deep learning reconstruction networks is constructed. At the encoder side, only four corner sub-aperture images (SAIs) are compressed with new CTU level bit allocation, a complete SAIs array is reconstructed by a LFI angular super-resolution network at the decoder side. To remove perceptual redundancy, we design a CTU level bit allocation strategy with the assumption of perceptual consistency, considering the characteristics of the human visual system in the bit allocation process. Experimental results show that for the proposed method with the designed CTU level bit allocation strategy, an average BD-BR savings of 13.676% in Y-PPSNR metric and 2.045% in VSI metric can be achieved. Compared with the high efficiency video coding (HEVC) intra coding model, the proposed method can achieve an average BD-BR savings of over 90%.