Retrieving Visual Concepts in Image Databases

This thesis addresses the problem of extracting and retrieving visual concepts in images databases, with the aim of developing a generic CBIR system applicable to a wide range of images. The system proposed in this thesis is scale and rotation invariant and easily updateable. The proposed approach is modular; the modules constituting the system are the following: • A visual attention module, that identifies pictorial subjects inside the image. This module is based on color, intensity and density of regions of different intensity inside the image. It is invariant under geometrical transformations (scale, rotation, etc) and robust to variation of mean image intensity. • A feature extraction module. The algorithm is applied to characterize regions identified by visual attention module as “salient”. It is resistant to noise, contrast changes and geometrical transformations. The features extracted by this method characterize both the color and the shape of the regions. • A supervised updateable neural structure, based on neural trees. This algorithm combines the advantages of neural and classical classifiers, allowing a fast update for example with new classes like in classical classifier, together with the supervised learning by examples of neural networks. • A semantic and geometric integration system, based on semantic networks, that integrates the data of the previous modules. The system allows to infer the presence of composite objects even if only one part of the object is correctly detected and classified Each of the first three modules represents an independent result in the field of visual attention, feature extraction and neural networks and its performances are evaluated separately. Alternative new original algorithms of visual attention and feature extraction developed during this thesis are proposed and their performances are compared with the ones selected for the whole system. All the proposed algorithms present good performances in comparison with the ones existing in literature. The whole system has been successfully applied to recognize pictorial subjects in a quite large domain (images of people and flowers available in internet). The philosophy associated with the proposed system, thanks to its modularity and to its updateability, provides a powerful method to implement efficient CBIR systems that can be easily adapted to the image domain addressed.