Indoor Scene Recognition Using Local Semantic Concepts

Abstract

This paper introduces a system for recognizing indoor scene images. The system aims at recognizing the environment illustrated in an image and assigning an appropriate semantic label to it. Developing such systems is one of the most important issues in the field of machine vision and robotics. They are extensively used in object recognition, image and video semantic recognition, motion detection, positioning, and robot direction. The overall algorithm involves three major steps. The first is to extract local information and spatial relationships in the images, modeling the environment based on those pieces of information, and using this model for the semantic sectioning of the image and labeling local areas using graph-based energy minimization algorithm. Results show that this system can perform as well as other object-based and 3D-image features-based environment recognition systems.