A Connectionist Technique for Data Smoothing

Filtering data to remove noise is an important operation in image processing. While linear filters are common, they have serious drawbacks since they cannot discriminate between large and small discontinuities. This is especially serious since large discontinuities are frequently important edges in the scene. However, if the smoothing action is reduced to preserve the large discontinuities, very little noise will be removed from the data. This paper discusses the parallel implementation of a connectionist network that attempts to smooth data without blurring edges. The network operates by iteratively minimizing a non-linear error measure which explicitly models image edges. We discuss the origin of the network and its simulation on an iPSC/2. We also discuss its performance versus the number of nodes, the SNR of the data, and compare its performance with a linear Gaussian filter and a median filter.