Research On Adaptive Weighted Binary Codes

With the development of internet and the wide application of multimedia technology,multimedia data such as image and video show a rapid growth trend,which also poses a severe challenge for multimedia retrieval technology.The hashing algorithms map the high-dimensional floating-point image features into a compact binary code,so that the nearest neighbor can be queried according to the hamming distance,which has the advantages of fast retrieval speed and small storage space.However,the current hashing technology still has the following problems:(1)when the encoding length is limited,the number of hamming distance values is small;(2)a large number of images in the retrieval results share the same hamming distance.In view of the above problems,the Data Statistics-based Query Adaptive Weighted Ranking Algorithm and the Product Quantization Data statistics-based Query adaptive weighted Reranking algorithm are proposed,and the main contributions are as follows:1.Data Statistics-based Query Adaptive Weighted Ranking Algorithm is proposed.To solve the number of hamming distance values is less when the coding length is limited,different weight values are assigned to different bits,so that the nearest neighbor can be queried according to the weighted hamming distance,and the distance value granularity is finer and the number is more.To make better use of the original feature information of the data set,the original feature statistical information of the data set is obtained before the binarization,and part of the information lost due to the binarization can be retained,and increase the distinction between the images.Because a large number of images share the same hamming distance,using data set binary codes and query vector binary codes to calculate the weight value,there will be a large number of samples share the same weight value,and the use of query vector original features is relatively low.Therefore,using query vector original features to calculate the weight value keeps the difference information of query vector,improves the adaptability of weight value,from query to the problem of sharing the same hamming distance for a large number of images is solved by using quantitative method.2.Product Quantization Data statistics-based Query adaptive weighted Reranking algorithm is proposed.To solve all binary codes share a set of weight values in each query,which leads to low accuracy of weight values,it is proposed to assign different weight values to the binary codes of different data sets,which can better distinguish the binary codes that share the same hamming distance,and achieve higher retrieval accuracy.To solve the problem of storage space,using the ideas of product quantitative,the data set is divided into several subspaces,within each subspace extraction data set of statistics,statistics through smaller dataset of Cartesian product for large data sets statistics,with less space to obtain larger statistics.The binary coding and query vectors are segmented into sub-binary coding and sub-query vectors by using the spatial segmentation method of data sets.Sub-weight values are constructed according to sub-binary coding,statistical information of sub-data sets and sub-query vectors.Segmentation of subspaces and calculation of sub-weights both increase the time consumption of the algorithm and increase the time complexity of the algorithm.Therefore,a smaller data clustering center is used to replace a larger data set sample,which can complete the acquisition of statistical information and weight value calculation in a short time,and improve the time efficiency of the algorithm.In GIST and CIFAR10 data sets,a comparison experiment of nearest-neighbor retrieval was set up.The experimental results show that the data statistics-based query adaptive weighted Reranking algorithm has excellent performance in nearest-neighbor retrieval.The product quantization data statistics-based query adaptive weighted Reranking algorithm effectively solves the problem of fewer weight values.Both of the two algorithms effectively improve the performance of neighbor retrieval and have better performance than the existing weighted algorithms.