Combining intra-image and inter-class semantics for consumer image retrieval

Unconstrained consumer photos pose great challenge for content-based image retrieval. Unlike professional images or domain-specific images, consumer photos vary significantly. More often than not, the objects in the photos are ill-posed, occluded, and cluttered with poor lighting, focus and exposure. In this paper, we propose a cascading framework for combining intra-image and inter-class similarities in image retrieval, motivated from probabilistic Bayesian principles. Support vector machines are employed to learn local view-based semantics based on just-in-time fusion of color and texture features. A new detection-driven block-based segmentation algorithm is designed to extract semantic features from images. The detection-based indexes also serve as input for support vector learning of image classifiers to generate class-relative indexes. During image retrieval, both intra-image and inter-class similarities are combined to rank images. Experiments using query-by-example on 2400 genuine heterogeneous consumer photos with 16 semantic queries show that the combined matching approach is better than matching with single index. It also outperformed the method of combining color and texture features by 55% in average precision.