Label relaxation using a linear mixture model

Abstract

A label relaxation method with a proportional estimation of mixed pixels (mixels) which is based on inversion problem solving techniques with a previously estimated proportion of neighbouring mixels is proposed. The method allows us to check the connectivity of separated road segments which arc observed frequently in the remote sensing satellite imagery. The experimental results with simulation data including observation noise show 73·5-98·8 per cent of improvement in terms of proportional estimation accuracy (RMS error), compared to the results from the previously proposed method with a generalized inverse matrix. Also the usefulness of the proposed label relaxation method with the proposed proportional estimation was confirmed for the investigation of the connectivity of the roads in the remote sensing satellite imagery.