A Multı-Instance Based Learning System for Scene Recognition

Abstract

Scene recognition is a frequently-studied topic of computer vision. The aim in scene recognition is to predict the general environment label of a given image. Various visual elements contribute to the characterization of a scene, such as its spatial layout, the associated object instances and their positions. In addition, due to the variations in photographic arrangements, similar scenes can be photographed from quite different angles. In order to capture such intrinsic characteristics, in this thesis, we introduce a multi-region classification approach for scene recognition. For this purpose, we first introduce a novel way of extracting large image regions, which are expected to be representative and possibly shared among the images of a scene. We utilize these candidate image regions within a multiple instance learning framework. In this way, we aim to capture the global structure of a given scene.