A hybrid tree approach for efficient image database retrieval with dynamic feedback

The need always exists for indexing mechanisms that can precisely retrieve imagery from a database, as well as maintain certain efficiencies for large-scale image database search. To achieve this, we developed a hybrid search tree called SKD-Metric tree. This novel approach merges the classification power of the statistical k-dimensional tree and the efficiency of computation of the Metric-tree for nearest neighbor (NN) search. Another feature of SKD-Metric tree is its flexibility to formulate a new metric function while the retrieval system utilizes user´s feedback to improve accuracy. In addition, unlike traditional relevance feedback approaches that, in most cases, sequentially search the entire database to obtain new retrieval results, SKD-Metric tree features a fast retrieval refinement procedure that needs to update only a small portion of the database. An extensive study, based on experiments performed for evaluating retrieval precision and computational efficiencies, is presented. We have applied our approach to a large-scale medical image database. The experimental results show that SKD-Metric tree can achieve a high accuracy rate with dynamic relevance feedback that requires much less computation than existing techniques.