Discovering recurrent image semantics from class discrimination

Abstract

Supervised statistical learning has become a critical means todesign and learn visual concepts (e.g., faces, foliage, buildings,etc.) in content-based indexing systems. The drawback of thisapproach is the need of manual labeling of regions. While severalautomatic image annotation methods proposed recently are verypromising, they usually rely on the availability and analysis ofassociated text descriptions. In this paper, we propose a hybridlearning framework to discover local semantic regions and generatetheir samples for training of local detectors with minimal humanintervention. A multiscale segmentation-free framework isproposed to embed the soft presence of discovered semantic regionsand local class patterns in an image independently for indexingand matching. Based on 2400 heterogeneous consumer images with 16semantic queries, both similarity matching based on individualindex and integrated similarity matching have outperformed afeature fusion approach by 26 and 37 in average precisions,respectively.