Research For All-purpose Image Retrieval System And High-dimensional Indexing

Today, the Content-Based Image Retrieval (CBIR) technique is beging widly used, However, most kinds of CBIR systems at present can only be applied in special field, all-purposed CBIR system is exiguous. Therefore, based on the existing technologies of domestic Dameng Database Management System, the design and development of a all-purpose CBIR system- DMIR has great significance.In the image search module of DMIR, a new high-dimentional indexing method named BDPlus method which is based on the anlysis of BID and iDistance method and suit for the K-nearest neighbor search is proposed. On the basis of the data distribution in high-dimensional spaces, BDPlus partitions the surface of dimensionality in a special way. By the definitions of bit code and transformation function, a high-dimensional vector can be first approximately represented and then transformed into a one-dimensional vector, the key managed by a B+-tree. In the process of k-nearest neighbor query, BDPlus get the Candidate result set by constructing a query scale and computing the intersection of query scale and space's partition. Using the difference between query node and partition which is computed from the bit-code weight, the intersection of query scale and space's partition can be shrink fastly during the k-nearest neighbor query process. Therefore, BDPus metheod has eximious filtering effect.The results of extensive experiments demonstrate that the BDPlus outperforms the existing index structures for KNN search in high-dimensional spaces. Test result of DMIR show that, by using the technology of dynamically loading feature extraction, feature matching and high-dimentional indexing algorithms, DMIR system can be applied in many fields of image retrieval requirements. The development of DMIR system lay the groundwork of the R & D for multimedia management technologies of domestic Dameng DBMS.