Dynamic properties of feed-forward neural networks and application in contrast enhancement for image

Abstract

This paper is concerned with three neurons feed-forward neural network model and more specifically with the study of dynamical behavior of the codimension one nilpotent singularity and 1:1 resonant Hopf bifurcation and outline possible image processing applications. Three neurons dynamical feed-forward neural networks use cross-coupling and feed-forward-coupling to form an nonlinear dynamic neural oscillator with the time delay. The theoretical basis of the pitchfork and 1:1 resonant Hopf bifurcation of feed-forward neural networks with delay is carried out and the analytical formulas are derived to define the various states of the system. The ultimate goal is to understand the dynamics and seek the application in image processing. It is shown that each of these states has a significant impact on the quality of the resulting image contrast enhancement. As application, aiming at the characteristics of remote sensing images with low-contrast and poor resolution textual information, an image enhancement method is presented. We show theoretically and numerically that the gray scale remote sensing image picture contrast is strongly enhanced even if this one is initially very small. The results show that the algorithm can significantly improve the visual impression of the image. Compared with the proposed algorithms in recent years, the information entropy are significantly improved.