A study on backpropagation networks for parameter estimation from grey-scale images

A large number of experiments have been done on the basic research of parameter estimation from images with neural networks. To obtain a better estimation accuracy of parameters and to decrease needed storage space and computation time, the architecture of networks, the effective learning rate and momentum, and the selection of training set are investigated. A comparison of network performance to that of the least squares estimator is made. The internal representations in trained networks, i.e. input-to-hidden weight maps or measuring models, which include statistical features of training images and have a clear physical and geometrical meaning, and the internal components of output parameters given by outputs of hidden neurons are presented.