Improved scene classification using region semantics and spatial context information

Abstract

In this paper, we present an improved scene classification algorithm to classify scene images using region semantics and spatial context information. Unlike existing solutions that rely on predefined prototypes and inflexible spatial context information, we propose using the region semantics dictionary to represent the scene conveniently, which can handle a variety of region semantics, and to explore more efficient spatial context information based on sparse representation characteristics. We first use statistical Gaussian mixing models (GMMs) to obtain region semantics and location information and effectively solve the parameters using an adaptive expectation-maximization (EM) algorithm, then establish a region semantic dictionary for scene interpretation. Moreover, we formulate the explored spatial context information as a convex optimization problem that can be solved using the inexact augmented Lagrange multiplier (IALM) method. Finally, we can predict the label information accurately according to Bayesian decision rule. Our extensive experiments have demonstrated that the improved algorithm can outperform several other previous methods on publicly available scene datasets; therefore, we also demonstrate and confirm the efficacy of handling errors due to variants of the local region semantics.