Image sorting and image classification: a global approach

A form sorting system designed for an industrial application is presented. This system has to deal with two main functions. On one hand it has to affect good labels to known form classes and on the other hand it has to reject unknown classes which can't be modelized. To achieve this objective a distinction between the task of classifying images and the one of sorting images is made. This lead to a system which structure is based on the combination of two classifiers. The first one is designed to affect a label to every pattern and the second one is designed to reject or not the proposition of the first one. Each classifier uses different primitives and different decision methods. By separating the objectives of both classifiers, we can define different performance evaluation criteria. In this paper we focus on the structure of the second classifier called the rejection unit and particularly on the methods used to determinate the rejection threshold. At last, results illustrating the different performance evaluation criteria are presented and discussed.