Learning as a supervised knowledge acquisition task

Models constitute representations of the real world, and thus, modelling implies the choice of a suitable set of parameters in order to build those representations. Obviously, the selection of the parameters will affect how well the models fit reality, and this becomes a central issue in any modelling process. Due to the complexity of some scenes, in our approach we make use of generic modelling to characterize every object in the scene, but include the possibility that every single object can be described by specific features, in order to facilitate later recognition processes.

Object modelling has been designed as a supervised task, where a teacher presents representative examples of objects in training images. Each real object is modelled as a set of object classes. An object class is defined as a prototype of a real object described in terms of colour and textural features under specific outdoor conditions. Conscious that not all outdoor conditions can be modelled for each one of the objects of interest, the system attempts to maximize outcomes from the available object knowledge. A proper combination of such classes will produce a suitable description of the object under any outdoor conditions.

Figure below details the interactive modelling process, which basically includes:

• A set of different images, corresponding to several outdoor conditions, is used to train the system. Three different attributes are used to define the outdoor conditions: (i) the season, (ii) the time of day, and (iii) the weather. With the aim of facilitating the learning task, these attributes can be initialized using linguistic values.
• Data acquisition. The teacher first selects meaningful regions representing object classes (i.e., real objects in specific outdoor conditions). Then, the learning application computes and registers a number of colour and texture features for each object class.
• Feature selection. The learning application selects a particular subset of features for characterizing each object class of interest.
• Learning verification. The learning application produces a visual feedback for the classes of interest.

• J.Freixenet. PhD. Thesis. Descripció d'Escenes Exteriors a partir d'un Aprenentatge Supervisat de les Característiques més Significatives dels Objectes. Supervisor: Dr. J. Martí. Presented in Girona, July 2000.
• X.Lladó, J.Freixenet, J.Martí. Selecció de Característiques mitjançant Algorismes Genètics. 3er Congrés Català d'Intel.Ligència Artificial (CCIA2000), pp. 152-156. Vilanova i la Geltrú, oct. 2000.
• J. Freixenet, J.Martí, X.Cufí, and X.Lladó. Use of Decision Trees in Colour Feature Selection. Application to Object Recognition in Outdoor Scenes. Proceedings of the IEEE International Conference On Image Processing (ICIP), vol. 3 pp. 496-499. Vancouver, Sept. 2000.
• X.Lladó, J.Martí, J.Freixenet, and Ll.Pacheco. A Novel Criterion Function in Feature Evaluation. Application to the Classification of Corks. Proceedings of the Quality, Automation and robotics Conference (Q&A-R 2000), vol. 2, pp. 275-280. Cluj-Napoca, Romania. May 2000.