Learning-based JND-directed SDR/HDR video prerprocessing scheme for perceptually lossless video compression

Abstract

This dissertation work aimed at targeting a preprocessing-based PVC scheme for Standard Dynamic Range (SDR)/High Dynamic Range (HDR) videos. It is a learning-based Just Noticeable Distortion (JND)-directed preprocessing scheme for perceptual video compression, which is called the SDR(HDR)-JNDNet. Our SDR(HDR)-JNDNet effectively suppresses the perceptual redundancy of SDR/HDR video signals so that the compression efficiency can be significantly enhanced by the following HEVC encoder. To our best knowledge, our work is the first approach to training a CNN-based model to directly generate the JND-directed suppressed frames of an SDR or 10-bit HDR video with the negligible perceptual quality difference between the decoded frames for the original input with and without the preprocessing by the SDR(HDR)-JNDNet. Also, we present a generative adversarial network (GAN)-based post-processing method for compression distortion reduction, which called SDR(HDR)-CARNet, optimized for the proposed preprocessing-based PVC scheme (SDR(HDR)-JNDNet) using GAN-based image restoration method. In extensive experiments, when the SDR(HDR)-JNDNet is applied as preprocessing for the SDR(HDR) video input before compression, it allows remarkably to save the required bitrates for SDR(HDR) test videos, with little subjective video quality degradation without increasing the computational complexity. In addition, when our preprocessing method (SDR(HDR)-JNDNet) and our post-processing method (SDR(HDR)-CARNet) are combined, we can obtain results of better perceptual visual quality while using fewer bitrates than using the conventional standard video codec for 4K-UHD/SDR and HDR videos. Our work is an essential study in real industries of broadcasting and video streaming companies because it has focused on high resolution and HDR images such as 4K-UHD/HDR, which requires a lot of bandwidth for transmitting.