Convolutional Neural Networks brought to the Computer Vision and Pattern Recognition fields a new way to characterize objects and scenes, achieving the state-of-the-art results on many classification tasks. Although these features encode highly discriminative characteristics in many classification domains, scene recognition still remains as a challenging task because of its particular attributes. In this work we present a new method to build a representative vector composed of global (e.g., environment’s structure) and local (e.g., characteristics of common objects of a particular scene) features from fragments of the scene by using robust sparse representations. The experiments have shown that the final resulting representation is more robust to noise and occlusion and it achieves the highest accuracy on well-known scene datasets, outperforming the previous results of the literature.

Prof. Erickson Rangel do Nascimento – Advisor (DCC – UFMG)
Prof. Anisio Mendes Lacerda (DECOM – CEFET)
Prof. Mario Fernando Montenegro Campos (DCC – UFMG)
Prof. William Robson Schwartz (DCC – UFMG)