Research On Locally Features Aggregating And Indexing Algorithm In Large-scale Image Retrieval

At present, increasing multimedia information resources, especially image data,are available due to the rapid development of the Internet and multimedia technology. How to implement the similarity retrieval based on content among image objects has become a very important research subject. Generally, similarity between image objects can be measured by distance between feature vectors. Due to the characteristic of high dimension, high-dimension indexing mechanism becomes the key technology for content-based retrieval in large–scale image database. To resolve the problem of'curse of dimension', organizing a large number of high-dimension feature vectors and designing efficient high-dimension indexing methods have become a great challenge for all the researchers.After analyzing the respective advantages and disadvantages of aggregation vector named Vectors of Locally Aggregated Descriptors (VLAD) of image and high-dimension indexing mechanism called Inverted File with Asymmetric Distance Computation (IVFADC) which combines product quantization, asymmetric distance computation and inverted index, new algorithms are proposed to improve the former shortcomings. As the relationship between local descriptors and clusters cannot be accurately described by hard-assignment in VLAD, more representative aggregation vector called Soft Assignment-Vectors of Locally Aggregated Descriptors (SA-VLAD) has been put forward through substituting soft-assignment for hard-assignment. Soft-assignment is a distributive strategy that the weight assigned to neighboring cells depends on the distance between the local descriptors and the cell centers. Besides, based on IVFADC, a new indexing scheme named Dispersed Assignment-Inverted File with Asymmetric Distance Computation (DA-IVFADC) is implemented through introducing dispersed-assignment in the stage of indexing to resolve the problem of massive distance computation and more search time brought by Inverted File with Asymmetric Distance Computation for increasing the number of inverted file lists to guarantee recall rate in the searching phase.Experiments show that compared with VLAD, SA-IVFADC can obtain higher accuracy rate of results in index mechanism when keeping storage space unchanged and inconspicuously increasing the search time. DA-IVFADC lessens the query time and improves the accuracy rate and recall rate at the expense of raising a handful of storage space. However, further analysis should be studied upon the computation of weight satisfying the spatial distribution of vectors and improvement of dynamics of index mechanism. Future work mainly focuses on these two aspects.