A method for still image interpretation relying an a multi-algorithms fusion scheme. Application to human face characterization

We propose a method for analysing still images relying on fusion of information obtained from several complementary processing channels. Each processing path works in three steps. A first stage, manipulating raw data, segments the image into aggregates. It is followed by a characterization stage that integrates information conveyed by aggregates. The terminal interpretation stage instanciates fuzzy rules. These rules constitute a loose model of objects present in the image. They semantically qualify aggregates, Once the interpretation has resumed on each separate channel, results of the different methods are collated on a fuzzy rule basis providing the final semantic identification of aggregates. This approach provides an interesting alternative for managing complementarity of several algorithms by introducing knowledge on each algorithm and on its intrinsic properties. The method is illustrated through an example of human face constituents characterization from black and white identity photographs. We give a thorough description of the method and provide quantitative results. These results show that a substantial improvement of recognition capabilities (50%) is obtained by fusing results of separate algorithms. (C) 1999 Elsevier Science B.V. All rights reserved.