The primary goal of the method described here is to allow reconstruction of images in high quality by using a universal image coder with easy bit-level access to MPEG-7-like low- and mid-level image features at the decoder. Natural images are mostly piecewise smooth. Therefore, the idea is to search for unsteady regions in the image, and to approximate the stationary regions separately, but smoothly combining both of them with great care.

For this, a sparse Mixture-of-Experts (SMoE) regression approach for encoding videos in the pixel domain is used, deferring drastically from the established DPCM/Transform coding philosophy. According to the invention, the MoE takes on the form of a Gaussian Mixture Regression (GMR) for multivariate nonlinear regression. The underlying stochastic process of the pixel amplitudes are modelled as a 3-dimensional and multi modal mixture of Gaussians with K modes. Therefore, each component in the MoE steers in the direction of the highest correlation. Experiments shows that – compared to JPEG – for a large class of images a considerable compression gain is achievable at low bitrates, while providing attractive low-level descriptors for the image. This way, the SMoE shows a strong resemblance to MoE neuronal networks, while providing a performance competitive with H.264.