Indexing pictures by key objects for large-scale image databases

This paper presents a method of indexing pictures by pattern recognition based on key image objects for large scale image databases. The idea is to generate an image ''signature'' for each database picture with respect to ''key'' objects. Such signatures are subsequently used to make picture retrieval more efficient by fast bitwise matching to reduce the search space before requiring compute-intensive similarity retrieval. Pictures with the same bit-pattern in their signatures are envisioned as a cluster. A set of key objects can be selected to ensure a good distribution of the pictures amongst the kernel clusters. We show that our mechanism of selecting keys is very effective in terms of minimizing subsequent similarity matching for both simulated and real image databases.