IMAGE CODING VIA PERCEPTUALLY INSPIRED GRAPH LEARNING

Most codec designs rely on the mean squared error (MSE) as a distortion metric, which simplifies parameter optimization but may fail to reflect perceptual quality. Alternative distortion metrics, such as the structural similarity index (SSIM), can be computed only pixel-wise, rendering them unsuitable for transform coding schemes. Recently, the irregularity-aware graph Fourier transform (IAGFT) emerged as a means to include pixel-wise information in the transform domain. This paper builds upon this idea by also learning a graph (and corresponding transform) for sets of blocks that share similar perceptual characteristics. We exploit that regions with different perceptual importance differ statistically to re-purpose graph learning algorithms considering perceptual criteria. We demonstrate the effectiveness of our method with both SSIM and saliency. We also propose a framework to derive separable transforms, including separable IAGFTs. An extensive empirical evaluation based on the 5th CLIC dataset supports the advantages of our approach.