NEURAL NETWORKS FOR IMAGE MODELING BY 2-DIMENSIONAL RANDOM-FIELDS WITH APPLICATION TO IMAGE COMPRESSION FOR TARGET ACQUISITION

The problem of target acquisition is considered to be a very involved process and is a serious challenge for the researchers. For several ap plications of target acquisition, it is worthwhile to compare the comp ressed and uncompressed images and the perceptual difference between t he two images is also significant. A new neural network technique of i mage modeling by 2-D random fields formulated in the form of an autore gressive moving average process driven by input white Gaussian noise w ith known statistics is presented. The proposed technique consists of two stages: (1) estimating the parameters of the model and (2) regener ation of the image with the knowledge of the model, its parameters, in itial conditions, and white noise. The problem of estimating the model parameters is formulated as an optimization problem solved by a singl e-layer neural network. Once the model parameters have been estimated as the adaptive weights of the network, the second stage reconstructs the picture from the model. This stage consists of recursively constru cting the image using the initial conditions of the original image, th e parameters of the model, and white Gaussian noise. Due to the adapti ve nature and the computational capability of the neural network, a hi gh-quality image is obtained with this approach. The proposed algorith m reduces the computational complexity and is recommended for the on-l ine image compression required in target-acquisition-type applications . As the image is constructed using fewer pixel values of the given im age in the form of initial conditions, and a few parameters of the mod el, very effective image compression is achieved. Several computer sim ulation examples are included to illustrate the effectiveness of the p roposed technique. (C) 1998 Society of Photo-Optical Instrumentation E ngineers.