A new scheme of merger information based on accuracy for image classification

Sébastien Desbouchages (1) , Noël Richard (1) , Anne-Sophie Capelle-Laizé (1)

1 SIC

In this paper, we introduce the use of accuracy criterion, estimated for each data sample in addition of the data feature. This complementary information is used during the combination of features to weight the data influence to produce a decision in a combination scheme. The major idea lies in the fact that all extracted measures or features are stained of errors, inaccuracy, and that kinds of errors must be taken into account during the combination and conflict management. We show in a generic case, how to express the accuracy criterion and how to use them in a belief function. In a classical classification problem, we proof also the contribution of the accuracy information in the improvement of the classification results.

Mots clés: Accuracy; Belief Function; Conflict; Dynamic Feature Selection

Domaines: Traitement du signal et de l'image [eess.SP]; Intelligence artificielle [cs.AI]; Vision par ordinateur et reconnaissance de formes [cs.CV]; Traitement du signal et de l'image [eess.SP]

Sébastien Desbouchages, Noël Richard, Anne-Sophie Capelle-Laizé. A new scheme of merger information based on accuracy for image classification. Workshop of The theory of belief functions, Apr 2010, Brest, France. p144. ⟨hal-00482275⟩