Abstract. This paper deals with supervised document image classification. An original distance based strategy allows automatic feature selection. The computation of a distance between an image to be classified and a class representative (point of view) allows to estimate a membership function for all classes. The choice of the best point of view performs the feature selection. This idea is used by an algorithm which iteratively filters the list of candidate classes. The training phase is performed by computing the distances between every class. Each iteration of the classification algorithm computes the distance d between the image to be classified and the chosen representative. The classes whose distance with this point of view differs from d are deleted in the list of candidate classes. This strategy is implemented as a module of A2IA FieldReader to identify the class of the processed document. Experimental results are presented and compared with results given by a knn classifier.

Abstract. This paper presents a semi-supervised document image classification system that aims to be integrated into a commercial document reading software. This system is asserted like an annotation help. From a set of unknown document images given by a human operator, the system computes regrouping hypothesis of same physical layout images and proposes them to the operator. Then he can correct them, validate them, keeping in mind that his objective is to have homogeneous groups of images. These groups will be used for the training of the supervised document image classifier. Our system contains N feature spaces and a metric function for each of them. These allow to compute the similarity between two points of the same space. After projecting each image in these N feature spaces, the system builds N hierarchical agglomerative classification trees (hac) corresponding to each feature space. The proposals for regroupings formulated by the various hac are confronted and merged. Results, evaluated by the number of corrections done by the operator are presented on different image sets.