Light field image coding: objective performance assessment of Lenslet and 4D LF data representations

Abstract

State-of-the-art light field (LF) image coding solutions, usually, rely in one of two LF data representation formats: Lenslet or 4D LF. While the Lenslet data representation is a more compact version of the LF, it requires additional camera metadata and processing steps prior to image rendering. On the contrary, 4D LF data, consisting of a stack of sub-aperture images, provides a more redundant representation requiring, however, minimal side information, thus facilitating image rendering. Recently, JPEG Pleno guidelines on objective evaluation of LF image coding defined a processing chain that allows to compare different 4D LF data codecs, aiming to facilitate codec assessment and benchmark. Thus, any codec that does not rely on the 4D LF representation needs to undergo additional processing steps to generate an output comparable to a reference 4D LF image. These additional processing steps may have impact on the quality of the reconstructed LF image, especially if color subsampling format and bit depth conversions have been performed. Consequently, the influence of these conversions needs to be carefully assessed as it may have a significant impact on a comparison between different LF codecs. Very few in-depth comparisons on the effects of using existing LF representation have been reported. Therefore, using the guidelines from JPEG Pleno, this paper presents an exhaustive comparative analysis of these two LF data representation formats in terms of LF image coding efficiency, considering different color subsampling formats and bit depths. These comparisons are performed by testing different processing chains to encode and decode the LF images. Experimental results have shown that, in terms of coding efficiency for different color subsampling formats, the Lenslet LF data representation is more efficient when using YUV 4:4:4 with 10 bit/sample, while the 4D LF data representation is more efficient when using YUV 4:2:0 with 8 bit/sample. The “best” LF data representation, in terms of coding efficiency, depends on several factors which are extensively analyzed in this paper, such as the objective metric that is used for comparison (e.g., average PSNR-Y or average PNSR-YUV), the type of LF content, as well as the color format. The maximum objective quality is also determined, by evaluating the influence of each block from each processing chain in the objective quality of the reconstructed LF image. Experimental results show that, when the 4D LF data representation is not used the maximum achieved objective quality is lower than 50 dB, in terms of average PSNR-YUV.