Classical object recognition methods from images usually rely on a statistical or structural description of the object content, summarizing different visual features such as contours, color or texture. Such features can then be used for the classification or the recognition of similar images. In this paper, we present a descriptor for complex objects represented in color images. The originality of this descriptor lies in an homogeneous representation of both shape attributes and spatial relations between structural regions extracted from objects using a robust decomposition method. We also propose different comparison strategies for this descriptor, based on specific matchings between structural regions of the objects. Results obtained on two datasets of color images suggest that the spatial relations between regions composing complex objects constitute interesting features for their description.