Large-scale Image Retrieval using Neural Net Descriptors

David Novak, Michal Batko, Pavel Zezula
2015 Proceedings of the 38th International ACM SIGIR Conference on Research and Development in Information Retrieval - SIGIR '15  
Keywords metric indexing; deep convolutional neural network; contentbased image retrieval; k-NN search ONLINE IMAGE RETRIEVAL SYSTEM One of current big challenges in computer science is development of data management and retrieval techniques that would keep pace with the evolution of contemporary data and with the growing expectations on data processing. Various digital images became a common part of both public and enterprise data collections and there is a natural requirement that the
more » ... l should consider more the actual visual content of the image data. In our demonstration, we aim at the task of retrieving images that are visually and semantically similar to a given example image; the system should be able to online evaluate k nearest neighbor queries within a collection containing tens of millions of images. The applicability of such a system would be, for instance, on stock photography sites, in e-shops searching in product photos, or in collections from a constrained Web image search. A successful content-based image retrieval system must stand on two pillars: effective image-processing technique to achieve high-quality retrieval, and efficient search techniques to make the system work in real time and on a large scale. Our demonstration uses the cutting edge approach of deep convolutional neural networks [4] to obtain powerful visual features that carry a certain semantic footprint of the image content. These descriptors compared by a distance function seem to very well correspond to the human perception of general visual similarity. Our main contribution is the search engine that can organize large volumes of these complex descriptors so that the similarity queries can be evaluated efficiently. The engine exploits our recent
doi:10.1145/2766462.2767868 dblp:conf/sigir/NovakBZ15 fatcat:ehofgbfl5ra6jagjbe52ytyojm