This paper has demonstrated the effectiveness of using features extracted from regions in segmented images. Radial Basis Function nodes have been employed to classify data points in a n-D feature space with centroids placed using limited training data using Kohonen's LVQ software. A manual classification of the most significant regions in an image database was conducted in order to assess the performance of this approach. Using ten resultant images, an average of 76.6% of the classified regions returned had a class which matched the key region used. This result was achieved with a minimal feature set, employing only size, position and 3 colour components. Work is currently under way to include texture and shape in the feature set which will greatly improve performance. This work also stresses the importance of exploiting user-feedback in database search systems of this kind.