On the use of SSIM in HEVC

Abstract

The Structural SIMilarity (SSIM) index has been attracting an increasing amount of attention recently in the video coding community as a perceptual criterion for testing and optimizing video codecs. Meanwhile, the arrival of the new MPEG-H/H.265 High Efficiency Video Coding (HEVC) standard creates new opportunities and challenges in perceptual video coding. In this paper, we first elaborate what are the attributes that make SSIM a good candidate for perception-based development of HEVC and future video coding standards for both testing and optimization purposes. We then address the computational issues in practical applications of SSIM in HEVC, in particular the trade-off between efficient computation and accurate estimation of SSIM when working with video codecs that have sophisticated block partitioning structures and aim for encoding videos with a wide range of spatial resolutions.

Content-weighted mean-squared error for quality assessment of compressed images

“…First, the design of our technique benefits from some principles. Image structure has an extremely important influence on the visual quality of human judgement, and it has been widely used in most existing IQA models [16][17][18][19][20][21][22][23][24][25][26][27][28] and some related applications [7][8][9][10][11][12][13]. In [29], the authors have pointed out that using a proper Gaussian kernel to filter the input original and distorted images can effectively extract image structure.…”

“…Moreover, it was found that, despite several uses of SSIM [9][10][11], in most cases MSE is employed for optimization in coding. "Why the state-of-theart FR IQA models cannot be applied to compression?"…”

“…the mean opinion score (MOS). Subjective quality scores are quite valuable, because on the one hand they can be used to justify the performance of objective IQA algorithms (e.g., LIVE [1], TID2008 [2], CSIQ [3], IVC [4], Toyama [5] and TID2013 [6]), and on the other hand, they are able to instruct denoising [7], restoration [8], coding [9][10][11], and compare tone-mapped operators [12][13][14]. But subjective methods easily suffer the drawbacks of being laborious, costly and time-consuming, and thus cannot be adapted to real-time applications.…”

On SSIM-bit rate comparison of HEVC encoders

“…In this paper, we attempt to address the main issues regarding the comparison of SSIM, which is a recommended quality criterion to be embedded into the video encoder [24]. First, by comparing the differences between the R-S and RD curves, we discuss why we need to develop a new comparison approach for the R-S curves.…”

“…By using the PSNR-based measure to evaluate the R-S performance, they employed an implicit assumption that the R-S curves had similar characteristics to the traditionally used RD curves (i.e., PSNR versus BR curves). However, the R-S curves have different shapes to the RD curves [24], and therefore the aforementioned BDPSNR & BDBR approach may not be an ideal scheme to calculate the perceptual improvement.…”