In order to browse a collection of images not only with textual query, a more complex model of the image is needed which takes into account for unstructured data. Object Oriented DBMS is suitable to deal with this data modeling ([4]). At the root of the OO paradigm is the concept of 'class', the software entity to represent real objects. A class is constituted by attributes specifying the features of the object, and by methods representing the rules to operate on attributes and acting as interface towards external enevironment. The attributes of a class can be represent, textual information, bit-mapped fields, abstract data types. Different degree of visibility can be established for the attributes in such a way that only authorized methods can access to them. Although Relational Database does not allow managing non-structured data they have the great advantage of the easiness of establishing relations. In order to hold this feature into OO-DBMS, two hybrid solutions as been proposed: the first one consists of modeling the relational database in the application classes. The data are stored into tables and when objects are created they retrieve the corresponding features from the database. Yet objects have to known how retrieve data from database. The second solution is to model the application in the database but many aspects of object-oriented paradigm are lost (inheritance, pointers, polymorphism can not expressed directly into table attributes). Both the above solution are inefficient and does not support direclty object models (OO-DBMS satisfy that condition). The problem of dealing with relations in OO-DBMS is solved by introducing special attributes (pointers) of objects. Some of these attributes can be classify as: optionality (present/not present), cardinality (1:1, m:1, m:n),  transferability (constant/variable),  semantics (is-part-of, is-set-of, is-in-hierarchy-with,..), similarity (color, texture, shape,spatial relations) ([7]).

     Then we propose an image model as follows:
 

     Each image object is implemented through a similar model (class model) (see figure 2).

     When an image is inserted into the database  objects are labeled according to a controlled vocabulary (UMLS), properties are identified such as geometry, hierarchical relations, functional relations, spatial relations. Moreover visual features are computed (object segmentation, contour extraction, color histogram, textures) and associations between formatted and unformatted data are established. let noice that unstructured data have meta and semantic attributes too. While some automatic procedures do exist to comput color histograms and textures, automatic pèrocessing for object segmentation, contour extraction is not yet automatic nor robust and manual involvement is strongly required. It is worth to point out that structure annotation by keywords is even an efficient indexing tool for objects in image and very efficient tools for a long time now have proved their performance ([12]). Yet visual browsing is an appealing feature and can not be excluded a priori from such a system which has to allow retrieval by textual keywords and attributes but also retrieval based on visual properties.

     Let focus on query operators. Due to typology of the attributes, different retrieval paradigms have to be allowed ([5]). They are based on color, texture, shape, spatial constraints, browsing, textual keywords and attributes. The retrieval based on visual properties rely with similarity attributes. With similarity an image is retrieved if is similar to a model or a scheme. Color similarity ([13]) is based on visual appearance defined by luminance and chrominance histograms, texture similarity is based on three main orthogonal components: bi-dimensional periodicity, mono-dimensional orientation, complexity obtained through Wold decomposition ([10]), shape similarity is based on geometrical properties of objects and spatial similarity is based on relative position relationship between objects into the image. This type of retrieval is not based on an exact matching between the attribute values used for querying the database and the retrieved data. When such a query is performed a set of possible matchings are recovered and some percentage of matching is assigned to everyone. Inside this set a finer search can be exploited. With the retrieval by browsing the user is allowed to navigate the database without any particular knowledge of the contained data. This service can extent from simple on-line help to driving the search process based on some attributes. Retrieval by attributes and keywords is very similar to the retrieval in standard DBMS. Query processing is here based on an exact matching. In figure 3 is shown the architecture of the proposed system.

    Let focus now with the problem of customize the retrieval based on user specialty. In biomedical field many attempts have been experimented in order to modeling users given their particular specialty through dedicated software agents ([2]). They rely on modeling a specialist when he interacts in a clinical environment with different issues and multimodal data. Nevertheless a few interest has been focused on modeling biomedical image user  (see figure 4).

     In figure 5  the agent architecture has been integrated into query processor. The scheduling is the following:
         1 .A user require a service to the system
         2. A software agent is built according to the adopted retrieval method and user skills
         3. The proper query operators are chosen
         4. A custom model of the image is also built according to user specialty
         5. Agent processes the query
         6. The coded result is sent to the interface to database block
         7. Mapping function recovers into the database a set relevant image
         8. Between this set, a degree of relevance to the user query is then computed
         9. Filtering function identifies primitive features to be retrieveb in the image according to image model.