A structured learning framework for content-based image indexing and visual query

Abstract

Nonspecific images in a broad domain remain a challenge for content-based image retrieval. As a typical example, consumer photos exhibit highly varied content, diverse resolutions, and inconsistent quality. The objects are usually ill-posed, occluded, and cluttered with poor lighting, focus, and exposure. Traditional image retrieval approaches face many obstacles such as semantic description of images, robust semantic object segmentation, small sampling problem, semantic gaps between low-level features and high-level semantics, etc. To manage the high diversity of images in a broad domain, we propose a structured learning framework to systematically design domain-relevant visual semantics, known as semantic support regions, to support index and query in a content-based image retrieval system. Semantic support regions are segmentation-free image regions that exhibit semantic meanings and that can be learned statistically to span a new indexing space. They are detected from image content, reconciled across multiple resolutions, and aggregated spatially to form local semantic histograms. The resulting compact and abstract representation can support both similarity-based query and compositional visual query efficiently. The query by spatial icons (QBSI) formulation is a unique visual query language to explicitly specify visual icons and spatial extents in a Boolean expression. For empirical evaluation, we perform the learning and indexing processes of 26 semantic support regions over 2400 heterogeneous consumer photos from a single family using Support Vector Machines. We report a $27\%$ improvement in average precision over a very high dimension feature-based approach on 24 semantic queries based on multiple examples and pooled ground truths. Last but not least, we demonstrate the usefulness of the visual query language with 15 QBSI queries that have attained high precision values at top retrieved images on the 2400 consumer images.