Fuzzy segmentation and sample-based object recognition in remote sensing images

Presents a method of segmentation, using fuzzy sets theory applied to digital images\´ analysis. Currently, pattern recognition methods are based on "segmentation-interpretation" cycles which have high processing time costs. The present approach allows one to convert these cycles into a linear application : only one segmentation and one interpretation. First, the authors apply the inverse of the gradient function in the raster image, to obtain a fuzzy set image. Using fuzzy sets theory they define a "convex combination of sets", also called random sets, which is composed of a set and a positive weight function linked to it. The membership function of a fuzzy set is obtained from its "convex combination of sets" representation. The fuzzy segmentation consists in extracting fuzzy regions in the fuzzy image. A fuzzy region is defined as a concentric set of ordinary and fuzzy regions. A crisp region is a set of connected pixels (with a non-zero membership value) obtained from a level-cut. In fact, the fuzzy segmentation splits the image into a tree of fuzzy regions which inherit characteristics of included ordinary regions (level-cuts). A particular object to be found (plane on an airport strip in an aerial image for example) may be defined by a sample image which is decomposed with the fuzzy segmentation algorithm into a hierarchical tree of characteristic regions. Then, the approach is to find the tree isomorphism (with a certainty degree) which links sample image fuzzy segmentation to fuzzy image in order to recognize this object.