A Granulometry Based Descriptor for Object Categorization

The progress in the area of object recognition in the last decade is impressive. The literature reports new descriptors, new strategies, new ways to combine descriptors and classifiers and new problems in a so fast pace that it is hard to follow the whole area. A recent problem in the area is the fine-grained categorization. In this work, to address this problem, we propose a descriptor based on the application of morphological granulometries in the map of edges of an image. This descriptor is used to characterize the distribution of lengths and orientations of edges and to build a model for generic objects. We also propose a new spatial quantization with an arbitrary number of levels and divisions in each level. This quantization is so flexible that adjacent regions may have overlapping areas to avoid breakages in the structures that are near the border of the regions as it happens in the traditional spatial pyramids. Both approaches are used in a challenging and recent object recognition problem, the categorization of very similar classes. The proposed descriptor was used along with other descriptors and the overall performance of our solution to this problem was about 8% better than other work using the bag-of-words approach reported in the literature. Our descriptor showed a result 12% better when compared to the results of other edge-related descriptor in the categorization of very similar classes.