A neural architecture for fully data driven edge-preserving image restoration

摘要: This paper proposes a neural architecture, based on two Hopfield nets interconnected with Boltzmann Machine, for completely data driven edge-preserving restoration of blurred and noisy images. Solving this problem entails the joint estimation image, degradation operator noise statistics, assuming that only are available. Since we consider class piecewise smooth images, modeled through coupled Markov Random Field an explicit, constrained line process, hyperparameters image model must be estimated as well. Adopting fully Bayesian approach, solution can obtained by maximization suitable distribution respect to field, hyperparameters, parameters. The very high computational complexity means in most practical cases it cannot applied, unless some approximations adopted. In paper, exploiting presence explicit binary propose which effective computing architecture interacting networks. particular, where main load is supported nets, one intensity other performing least square blur coefficients. Machine used following modalities: running learning. modality, updates process; learning performs ML interpreted weights cliques neurons. Simulation results provided highlight feasibility efficiency adopted methodology.