Content based retrieval of images for planetary exploration

Planetary missions generate a large quantity of image data. In flight operations or on servers such as Planetary Data Systems (PDS) these data products are only searchable by keys such as the Sol, spacecraft clock, or rover motion counter index, with little connection to the semantic content of the images. During mission science operations the science team typically pores over each and every data product returned, which may not require more sophisticated organization and search tools. However, for analyzing existing image databases with thousands or millions of images, manual searching, matching, or classification is intractable. In this paper, we present a method for matching images based on similarities in visual texture. For every image in a database, a series of filters are used to compute the response to localized frequencies and orientations. Filter responses are turned into a low dimensional descriptor vector, generating a 37 dimensional fingerprint. At query time, fingerprints are quickly matched to find images with similar appearance. Image databases containing several thousand images are pre-processed offline in a matter of hours. Image matches from the database are found in a matter of seconds. We have demonstrated this image matching technique using three sources of data and carried out user tests to evaluate matching performance by hand labeling results. User tests verify approximately 40 % false positive rate within the top 14 matches. This represents a powerful search tool for databases of thousands of images where the a priori match probability for an image might be less than 1%.