Eugen Barbu

Automatic Ground-truth Generation for Document Image Analysis and Understanding

Performance evaluation for document image analysis and understanding is a recurring problem. Many ground- truthed document image databases are now used to evaluate algorithms, but these databases are less useful for the design of a complete system in a precise context. This paper proposes an approach for the automatic generation of synthesised document images and associated ground-truth information based on a derivation of publishing tools. An implementation of this approach illustrates the richness of the produced information.

Using bags of symbols for automatic indexing of graphical document image databases

A database is only usefull if it is associated a set of procedures allowing to retrieve relevant elements for the users’ needs. A lot of IR techniques have been developed for automatic indexing and retrieval in document databases. Most of these use indexes depending on the textual content of documents, and very few are able to handle graphical or image content without human annotation. This paper describes an approach similar to the bag of words technique for automatic indexing of graphical document image databases and different ways to consequently query these databases. In an unsupervised manner, this approach proposes a set of automatically discovered symbols that can be combined with logical operators to build queries.

Using texture-based symbolic features for medical image representation

At present time the Internet has become a major source of information and a powerful didactic tool. Furthermore, the development of digital equipment, allows to acquire and store large quantities of medical data, including images. In the context of the CISMeF on-line health-catalogue, our work is centered on the automatic categorization of medical images according to their visual content, for further indexation and retrieval tasks. The aim of the present study is to assess the performance of a new image symbolic descriptor for medical modality, anatomic region and view angle image categorization. This descriptor is issued from the unsupervised partition of statistical and texture image sub-block representations. A medical image database of 10322 images from 33 classes was ground-truthed by a domain expert. Despite the complexity and variability of medical images, the compact symbolic representation approach proposed in this paper achieves high recognition rates. Thus, using kNN classifiers, we obtain an average precision of 83% and a top performance of 91.19%

Graph Classification Using Genetic Algorithm and Graph Probing Application to Symbol Recognition

We present in this paper a graph classification approach using genetic algorithm and a fast dissimilarity measure between graphs called graph probing. The approach consists in the learning of a set of synthetic graph prototypes which are used for a 1NN classification step. Some experiments are performed on real data sets, representing 10 symbols. These tests demonstrate the interest to produce prototypes instead of finding representatives which simply belong to the data set

MedIC at ImageCLEF 2006: automatic image categorization and annotation using combined visual representations

The CISMeF group participated at the automatic annotation task of the 2006 ImageCLEF cross-language image retrieval track, employing the MedIC module. The module is designed to automatically extract annotations using image categorization. For the 2006 ImageCLEF annotation experiments we used two sets of visual representations: the first based on the PCA transformation of combined textural and statistic low-level visual features and the second oriented towards compact symbolic image descriptors extracted from the same low-level features. Only the first set of features was present in the actual competition and obtained the fourth rank, with only 1% of error more than the most accurate run of the competition. The comparison with the second representation set was conducted after the official benchmark, on the validation data set, and obtained similar results, but with significantly smaller image signatures.