BAYESIAN ENHANCEMENT OF A SVM BASED IMAGE SEARCH ENGINE

Nowadays the imagery satellites in orbit acquire a huge volume of data. These data need to be stored and pro- cessed to extract the information within each image in order to index it or to interpret it. An important aspect in an indexation system of a large satellite image database is finding relevant information for specific purposes during query sessions. For this pur- pose Content Based Image Retrieval (CBIR) systems are employed. These systems are trying, based on a simi- larity measure, to retrieve relevant information from the database. Since the retrieval precision is not high, the human decision is introduced in the processing loop by means of a Human Machine Communication (HMC). This consist in a dialog session between the human com- ponent and the machine based on giving relevant / irrele- vant examples. This is the case of Relevance Feedback (RF) algorithms. They boost the precision of the retrieval process by in- troducing in the process the human decision. The human decision is represented by the indication during the hu- man - machine dialog of examples within the database as "relevant" or "irrelevant". Thus the system learns the dis- criminant properties of the relevant samples and extends the classification to non-annotated samples.