Statistical-mechanical approach for analog neural network model used in image restoration

摘要

The ability to restore an image from signals received through a noisy channel is an important concern. This issue is related to the physics theory of spin-glass. In the theory, the Ising spin system is usually used for image restoration; however, a lot of calculation time is needed to obtain precise solution. As a result many researchers substitute the Ising spin model with the analog neural network model. We analyzed the analog neural network ability applied to the image restoration problem using the mean field theory. With the conventional image restoration method, the estimated overlap with the analog neural network model is equivalent to that of the Ising spin model. When parity codes are sent, the analog neural network's ability does not improve over the Ising spin model. If the noise variable is small, however, the performance of the analog neural network model is as good as the Ising spin model.