[Paper] Blind PSNR Estimation of Compressed Video Sequences Supported by Machine Learning

Abstract

The peak signal-to-noise ratio (PSNR) used as an index of image quality usually requires original images, but this is difficult for consumer generated content such as videos on YouTube. Therefore, we developed two blind PSNR estimation methods without bit-stream analysis in which multiple support vector machines are prepared to learn differently encoded images in PSNR; using an entire frame and dividing the frame into two areas. We confirmed that higher estimation accuracy is possible for the latter method against that using the entire frame.