As the growth of digital image collections continues so does the need for efficient content based searching of images capable of providing quality results within a search time that is acceptable to users who have grown used text search engine performance. Some existing techniques, whilst being capable of providing relevant results to a user's query will not scale up to very large image collections, the order of which will be in the millions. In this paper we propose a technique that uses text based IR methods for indexing MPEG- 7 visual features (from the MPEG-7 XM) to perform rapid subset selection within large image collections. Our test collection consists of 750,000 images crawled from the SPIRIT collection (discussed in section 3) and a separate set of 1000 query images also from the SPIRIT collection. An initial experiment is presented to measure the accuracy of the subset generated for each query image by taking the top 100 results of the subset, and comparing those to the top 100 results derived from a complete ranking of the collection for that query image. Ranking is performed via L2 Minkowsky distance measures for both sets.