Image Labeling by Integrating Local, Middle and Global Information

Abstract

We carry out image labeling based on probabilistic integration of local, middle and global information. Local information is effective for capturing color and texture pattern. Middle information is obtained from patches which are larger than local regions and is able to incorporate context information. Global information obtained from an entire image helps to decide the presence of categories in the scene. In the experiments using the MSRC21 dataset, labeling accuracies are much improved by integrating local, middle and global information. Our method gave the state-of-the-art performance.