Identification of Choquet integral’s parameters based on relative entropy and applied to classification of tomographic images

S. Jullien, G. Mauris, L. Valet, P. Bolon, S. Teyssier.

A method based on relative entropy is proposed to identify Choquet integral parameters for classification purposes. While the entropy-based method initially developed by Kojadinovic [11] assigns a high importance to the attributes which allow to differentiate many classes, the relative entropybased method proposed focuses on the researched classes. This is done in order to obtain the attribute confidence by the integration of expert knowledge concerning these classes. The values of the relative entropy for different subsets of attributes are considered as the values of Choquet capacity coefficients. The proposed identification method is integrated in an aiding system aiming at interpreting 3D-tomographic images of electrotechnical parts made of composite materials and manufactured by Schneider Electric. Four attributes extracted from the tomographic images are aggregated by the Choquet integral using the parameters identified by the proposed method. The result is a 3D cartography of the sought-after regions within the tomographic image. Quantitative assessments of classification highlight the relevance of the proposed approach.

PDF full paper