Image segmentation by successive background extraction

摘要

An adaptive thresholding method that extracts the image background in two phases is introduced. First, statistical parameters are measured in rectangular image regions. Background-homogeneous regions are detected with criteria based on statistical theory. Information from background-homogeneous regions is propagated to other regions. The background's graytone distribution of each region is extracted from measurements or estimates of the statistical parameters. The second phase of the method refines background extraction in non-homogeneous regions with actual statistical parameters if possible, else with rules that predict their values. Pixels are then classified into multiple labels corresponding to brightness relative to the local background. Good results have been obtained in industrial and outdoor scenes of moderate complexity.