No-Reference Metric Design With Machine Learning for Local Video Compression Artifact Level

In decoded digital video, the local perceptual compression artifact level depends on the global compression ratio and the local video content. In this paper, we show how to build a highly relevant metric for video compression artifacts using supervised learning. To obtain the ground truth for training, we first build a reference metric for local estimation of the artifact level, which is robust to scaling and sensitive to all types of compression artifacts. Next, we design a large feature set and use AdaBoost to create no-reference metrics trained with the output of the reference metric. Two separate trained no-reference metrics, one for flat and one for detailed areas, respectively, are necessary to cover all types of artifacts. The relevance of these metrics is validated in a compression artifact reduction application, using objective scores like PSNR and BIM, but also a subjective evaluation as proof. We conclude that our created reference metric is an accurate local estimator of the compression artifact level. We were able to copy the performance to two no-reference metrics, based on a weighted mixture of low-level features. Our new metrics enable a far superior performance of artifact reduction compared to relevant alternative proposals.