Prototypicality effects in global semantic description of objects. (arXiv:1801.04331v1 [cs.CV])

We propose a new approach to face the semantic features descriptions of objects based in the prototypicality effects of prototypes theory. Our descriptor, called global semantic descriptor, is capable of coding and storing a semantic (central and peripheral) meaning of object. Our model compute the semantic prototype of object using features extracted by CNN-classifications models. We propose a simple method to reduce the dimensionality of semantic prototype without semantic information lost. We demonstrated that our descriptor preserved de semantic information used by the CNN-models for classifications task. The experiments on MNIST and ImageNet Datasets demonstrated that our model allows clustering the elements into family resemblance (typicality value) within the category.

from cs updates on arXiv.org http://ift.tt/2mF9kM4
//