CMLBPIncoherent: a New Contextual Image Descriptor for Scene Classification

Abstract

Scene classification has been widely studied in the last years and is a significant computer vision problem. Currently, this application has been involved by several researchers, since the creation of robust datasets until the utilization of high-level techniques with great performance. However, due to the high complexity and variability from scenes, there is still so much to be studied and implemented aiming to improve better results. This paper proposes a new visual descriptor for scene classification, the CMLBPIncoherent, which models the distribution of local structures through Local Binary Pattern (LBP) algorithm combined with contextual information, and discards homogeneous regions by Color Coherent Vector (CCV) algorithm. The proposal was tested using some traditional scene databases and the results illustrate a better performance in comparison with other visual descriptors known in the literature.