Bootstrap Dual Complementary Hashing With Semi-supervised Re-ranking On Large Scale Image Retrieval Problem

With the rapidly development of multimedia in the internet,the image retrieval problem attracts more and more attention around the world.Given a query image,the retrieval problem is to return the similar images as retrieval result.The image retrieval methods should have excellent performance on speed,precision and recall.The hashing-based image retrieval methods show their great performance on image retrieval problem for search efficiency.Hashing-based methods project images into hashing space and use the hamming distance to measure the similarity between images.For the hashing-based methods,each image usually occupies several bytes of storage and the query process is very efficient.Excellent hashing method should ensure that the similar images have close hamming distance and dissimilar images have large hamming distance.Current hashing methods can be categorized into supervised,semi-supervised,and unsupervised according to the requirements of label information.The supervised hashing methods need fully labeled dataset,and the unsupervised hashing methods ignore the label information of dataset.Actually,the dataset usually provides partially label information;the hashing methods should use the label information to improve performance.Semi-supervised hashing methods require only partly labeled dataset,guaranteeing the performance and more actual in the same time.The multi-hashing methods have excellent performance on the image retrieval problems.DCH trains sequentially.But DCH discards correctly hashed images for the training of new hash tables and the training of hash functions considers only the error of the previous function.The bootstrap dual complementary hashing with semi-supervised re-ranking on large scale image retrieval problem(BDCHR)is proposed based on the drawback of DCH.BDCHR is a semi-supervised multi-hashing method.BDCHR trains hash tables by increasing the weight of error mapping images and trains hash functions considering the error of all the learned hash functions in this table.Moreover,the performance of each hash function of each hash table is usually different.The BDCHR propose the semi-supervised re-ranking to improve the retrieval performance.