Semantic region labelling using a point pattern analysis

Several recent studies and researches focused on the combination of global and fine local region description using points of interest. The main benefit of global approach is that homogenous parts of the image can easily be described by means of global attributes whereas small details are ignored. In the other hand, as the region gets smaller and with high photometric variability, points of interest are more appropriate to carry local description. In this paper, we propose a new semantic labelling of regions using their interest point spatial dispersion. We introduce a point-based criterion to label regions into homogeneous and textured classes. Our point based criterion is based on a point pattern analysis study and has been validated on a multispectral satellite image database. In our work, we combine the region and point description by constructing a descriptor provided by a new semantic labelling of regions for boosting the object recognition.