Abstract
In recent years, rate-distortion optimization for structural similarity (SSIM) has been well developed and studied to improve visual quality for video coding. SSIM is also well matched to human visual system and has better perceptual quality than MSE. Because of complex computation of SSIM, we used Noise-to-Signal Ratio (NSR) to approximate SSIM. First, in order to explain that NSR has good correlation with SSIM, we did some analysis on N SR and SSIM. What is different from previous work is that we proposed four-8x8-macroblock and nine-8x8-macroblock as our calculation unit instead of one-16x16-macroblock. Then we compared these three configurations and chose the better one: four-8x8-macroblock. From the experiments above, we knew NSR has good correlation with SSIM indeed, so we used NSR as our distortion metric. We also found that regions with different variances can use different quantization steps to improve quality in terms of SSIM. Therefore, we proposed a variance-adaptive quantization algorithm which uses variance to scale quantization parameters. The experiment results using JM17.0 shows that we can save 12% - 23.3% bitrate compared with the original JM17.0.