Image restoration can be regarded as a minimization problem

which can be solved by Boltzmann machine

since it is certainly fall into global minimum not local minimum by its inherent random neural network. Following improvements are done on traditional Boltzmann Machine: firstly integrate Paik into Boltzmann Machine; secondly extend serial model to parallel model; thirdly subunit step is adopted to increase precision; finally make the step adaptive to trade off the contradiction of convergence precision and convergence velocity. Each improvement was expatiated via theories

convergence analysis

and residual. Experiments demonstrate that this proposed method can approximate the overall minimum of the energy function infinitely

and the restored image by this method is superior to that of traditional Boltzmann.