Distributional semantic content-based image retrieval

This paper introduces a novel approach for content-based image retrieval called "distributional semantic content-based image retrieval", where the image semantic is computed from a co-occurrence matrix derived from a collection of candidate images. Near synonyms or related visual contents have just similar contexts. The candidate and query images are represented as vectors in the multi-dimensional space. Then the cosine similarity function is used to characterize the semantic proximity between image vectors. The prototype and preliminary experiments demonstrate the indexing visual contents of images on the basis of their context. The images in collections could be retrieved without difficulty even if a query is composed of related visual contents.