Abstract
We introduce the multivariate fields of experts, a new framework for thelearning of image priors. Our model generalizes existing fields of expertsmethods by incorporating multivariate potential functions constructed viaMoreau envelopes of the $\ell_\infty$-norm. We demonstrate the effectiveness ofour proposal across a range of inverse problems that include image denoising,deblurring, compressed-sensing magnetic-resonance imaging, and computedtomography. The proposed approach outperforms comparable univariate models andachieves performance close to that of deep-learning-based regularizers whilebeing significantly faster, requiring fewer parameters, and being trained onsubstantially fewer data. In addition, our model retains a relatively highlevel of interpretability due to its structured design.