Video Compression Artifact Suppression Using Perceptual Criteria

Abstract

This thesis explores the application of perceptually motivated optimization techniques in training neural networks for suppressing video compression artifacts. These techniques aim to align optimization criteria with the Human Visual System (HVS), a historically challenging task due to the complexity of modeling perceptual quality. However, the advent of neural networks has made it increasingly feasible to incorporate perceptual considerations into video processing.

The research introduces two primary approaches that leverage perceptual criteria to enhance the visual quality of compressed video. The first approach focuses on creating a differentiable approximation of the Video Multimethod Assessment Function (VMAF), a widely used yet non-differentiable perceptual video quality metric. We call this model ProxVQM. ProxVQM improves upon existing models by incorporating temporal features through the use of multiple frames, rather than a single frame. This improvement reduces the maximum error from +/-16 VMAF when using a single-frame model to +/-8 VMAF with the multi-frame approach. By incorporating ProxVQM into a restoration network as a loss function we then directly integrate perceptual quality into the optimization process. Training with this approach leads to a gain of up to +4 VMAF when compared to training using a standard pixel loss alone.

The second approach explores the development of a critic network within a Generative Adversarial Network (GAN) framework, with modifications to ensure that the critic is perceptually relevant. This GAN-based method employs deep features from pre-trained networks to create a perceptual critic, which is then used to adversarially train a video compression artifact suppression network, resulting in significant enhancements in perceptual video quality. Training in this manner demonstrates significant improvement in deep feature quality metrics (10% improvement in KID and 16% improvement in DISTS) when compared to training with traditional critic architectures. This is further underscored through the use of a subjective study which shows a preference of our proposed method compared to other state of the art approaches by 5 MOS on a 100 MOS scale.

The thesis also explores the practical application of these techniques within a real-world video transcoding pipeline, demonstrating their effectiveness in meeting target bitrate constraints while maintaining high perceptual quality. These methods were submitted to the video compression track of the Challenge on Learned Image Compression (CLIC) 2024, where our approach placed third overall. This achievement is particularly encouraging, as our method utilizes a neural post-processor alongside a current-generation codec, while the top two submissions employed next-generation codecs with in-loop neural network filters, which are known for their superior performance. The CLIC challenge is evaluated through a large-scale subjective study, further highlighting the ability of our model to produce perceptually relevant results.

The key contributions of this work include the development of a CNN architecture for differentiable perceptual quality assessment, the integration of perceptual loss functions into video processing pipelines, and the design of a GAN with a perceptual critic tailored for compressed video enhancement.