Sparse coding for histograms of local binary patterns applied for image categorization: Toward a Bag-of-Scenes analysis

In this work
¹
, we propose a novel approach for image categorization, which we will refer to as Bag-of-Scenes (BoS). It is based on the association of Sparse coding (Sc) and pooling techniques applied to histograms of multi-scale Local Binary Patterns (LBP) and its improved variant. This approach can be considered as a 2-layer hierarchical architecture. The first layer, encodes general local patch's structure via histograms of LBP, and the second, encodes the relationships between pre-analyzed LBP-scenes. Our method outperforms SIFT-based approaches using Sc techniques and can be trained efficiently with a simple linear SVM. Our BoS method achieves 87.02%, 87.71% and 79.05% of accuracy for Scene-15, UIUC-Sport and Caltech101 datasets respectively.