A FDO Neural Network Model and Its Application to Image Recognition

A full domain optimum neural network (FDONN) and its application to image recognition are proposed in this paper. In general, we cann't ensure the devised neural network to converge to a global minimum. In this paper, we use a method of devising stable points firstly and basins of attraction laterly. This method increases speed and correctness of recognition. Owing to the nature that the dimensions of image are very large, we present a invarint transformation to decrease the dimension of image and not to change the distance between them. In this paper, we prove the properties of the neural network: quality of convergence, speed of convergence and invariance of mapping. Several computer simulation examples involving trained and recognized targets are given to illustrate the usefullness of our method. The recognitized targets include noise-added targets and targets cut off a little part.