An Unsupervised Classification Framework for Content Based Image Retrieval using Semantic Assistant Based Visual Hashing

To provide high quality content-based search services over huge volume of image collections, both efficiency and effectiveness are important issues. Advanced indexing structure is essential to scale the big data space and facilitate accurate search. The most naive approach for CBIR is to sequentially compare query image with each sample stored in the database. Its linear complexity leads to the poor efficiency and low scalability in real environment. Also, visual features usually have high dimensions. How to solve the curse of dimensionality is still an open research question, which has not been addressed properly. In most real CBIR applications, approximate retrieval results can sufficiently satisfy user’s information needs. An rising technology to support scalable content-based image retrieval (CBIR), hashing has been recently received great attention and became a very active study domain. In this study, to propose a novel unsupervised visual hashing method is known as semantic-assisted visual hashing (SAVH). Renowned from semi-supervised and supervised visual hashing, its core design is to effectively extract the rich semantics latently embedded in auxiliary texts of images to enhance the usefulness of visual hashing without any explicit semantic labels. To achieve the target, a unified unsupervised skeleton is developed to learn hash codes by concurrently preserving visual similarities of images, integrating the semantic assistance from assisting texts on modeling high-order relationships of inter-images and characterizing the correlations linking images and documents.