This technology is a system for performing visual searches from a mobile device in wireless communication with a server. Local visual features are extracted from a query image acquired by the mobile device and encoded using a compact hash method. The encoded features and image boundary information are transmitted to a server that matches the features against a database. Geometry verification is performed on the match results to determine whether the query image may be a geometric transformation of an image in the database. The query image's boundary data is used to rank the search results before they are transmitted back to the mobile device for viewing by the device user.
Mobile visual search systems typically must extract relevant features of a captured image and transmit them to a server that attempts to match them over a database. Since transmission of visual features in graphical format can consume considerable bandwidth, there is a need to compress extracted image features. As the use of compressed graphics formats for this purpose does not afford a sufficient reduction in bandwidth usage, features must be encoding using some other method. By employing a bag of hash bits method to encode local images features, the technology can represent features using only a small number of bits. The hash bit representation also accelerates search speeds on the server by enabling flexible creation of the data indexing structure used to match encoded features. Search accuracy is boosted by using image boundaries to rerank the returned search results.
The technology's ability to outperform existing state-of-the-art mobile search methods was demonstrated over multiple sets of product images each containing between 300,000 and 400,000 images.
Tech Ventures Reference: IR CU12200