Ploughing through digital image collections
Not so long ago, there was one sole option allowed for searching images: keywords. But, things have drastically changed over the past years with image recognition technology. Any given image can be converted into an array of computed vectors, describing in mathematical terms the visual content. Dedicated search engines leverage image recognition technology to allow users to track copies of a specific image as well as modified versions. The VIAMASS (Visual recognition made super-scalable) project team is credited with innovative methods that have drastically increased the performance of such visual queries. However, researchers saw another challenge in automatically uncovering visual links between images. Before VIAMASS, there was no methodology available for efficiently and accurately discovering such visual links between images within a very large collection. Previous efforts resulted in algorithms able to detect frequent visual patterns while rare matches were left undetected. Researchers started with the development of radically new image representations that would allow them to tackle the visual recognition issues they aimed to solve. The next step was to find algorithmic solutions that proved capable of finding subsets of vectors representing an identical object in different images. Lastly, they proposed new coding methods to compare such sets of vectors in a huge image collection. The five-year research effort illustrated the merits of several unconventional approaches to similarity searches through demonstrators featuring selected visual links in images. VIAMASS work will pave the way to better representations for the conventional query by sample, impacting the processing chain of visual links searches.
Keywords
Digital image, image collections, visual links, image recognition technology, VIAMASS