Large-Scale Discovery of Spatially Related Images

Reviewer: Dick Brodine

How do you retrieve all of the images in an image database that are similar to an image that you have chosen__?__ To date, most retrieval algorithms gather images that are labeled with identical text tags. Chum and Matas' paper presents an algorithm that matches images on the basis of image content. Images are summarized by an image descriptor vector, and mini-hash, a hashing algorithm, is applied to each descriptor. Descriptors with identical hash values are placed in the same bin. Similarity is calculated for each combination of descriptors in a bin. Image pairs with similarities greater than a certain level (threshold) are forwarded to a spatial consistency test. The spatial consistency test performs a comparison of the image feature geometries. Image pairs that pass spatial consistency are classified as cluster seeds. The cluster seeds are employed to amass other images that resemble the seed into the cluster, thereby grouping similar images. When the user supplies an image as part of a query, that image is compared to each of the known clusters. If a match is found, the images in that cluster are returned to the user. This paper should be of interest to researchers in the field of image processing. It assumes a fair knowledge of statistical image representation-that is, visual words-on the part of the reader. That being said, the paper cites a number of references on this concept and other related topics. Online Computing Reviews Service