Recognizing architecture styles by hierarchical sparse coding of blocklets

Abstract

In this work, we propose a novel architecture style recognition model by introducing blocklets that capture the morphological characteristics of buildings. First, we decompose a building image into a collection of blocks, each representing a basic architecture component such as a stone pillar. To exploit the spatial correlations among blocks, we obtain locklets by extracting spatially adjacent blocks, and further formulate architecture style recognition as matching between blocklets extracted from different buildings. Toward an efficient blocklet-to-blocklet matching, a hierarchical sparse coding algorithm is proposed to represent each blocklet by a linear combination of basis blocklets. On the other hand, toward an effective matching process, an LDA [25,1]-like scheme is adopted to select the blocklets with high discrimination. Finally, we carry out architecture style recognition based on the selected highly discriminative blocklets. Experimental results on our own compiled data set demonstrate that the proposed approach outperforms several state-of-the-art place/building recognition models.