Discovering latent spatial structured patterns using graph models for scene classification

Abstract

Despite progress in scene recognition tasks such as image classification and attribute detection, computers still be difficult to understand the scenes as a whole. Existing methods often ignore global spatial constructed pattern among different local semantic objects. This paper propose a method for discovering the Latent spatial structured patterns to describe the visual semantic characters of images to improve the performance of scene recognition tasks. Unlike the existing approaches that mainly rely on the discriminant visual feature cues, we learn the latent spatial structured pattern to model the interaction relationships by using the graph models, which consider semantics and their localization information. We first train the pLSA models to obtain the latent semantic topics. Then we construct the graph models to discover the latent spatial structure patterns with combing the character vector and localization cues. Meanwhile, we treat the edge in model as link-affinity matrix to describe the interaction relationships between semantics. The extensive experiments on public datasets have demonstrated that the suggested method can significantly boost the performance of scene classification tasks.