Research And Application Of Hashing-based Binary Codes

With the rapid development of various image processing tools and shared Inter-net,efficient similarity search in large image databases is becoming a significant chal-lenge in many computer vision applications.Searching nearest neighbors of a query by scanning all the data in a large database has a linear time complexity,which is very inefficient and expensive.Fortunately in many applications,it is acceptably suf-ficient to return approximated nearest neighbors(ANN)instead of the exact ones.Bi-nary code learning techniques have been actively studied for hashing based nearest neighbor search in recent years.Hashing-based binary codes learning techniques have been proposed and widely studied for ANN in recent years.We focus on the study of hashing-based binary codes learning,and further proposed solutions.And we apply binary codes learning techniques to Bag-of-Feature image retrieval.Firstly,we propose a novel binary codes learning method based on feature recon-struction.Most existing techniques directly map the data into a Hamming space,which ignores the inherent property that original features may lie in different subspaces.Our method obtain latent components via feature decomposition which is based on decon-volution.This approach distinguishes the original data better and consequently im-proves the performance of binary codes.As the experiments showed,the algorithm is able to enhance the discriminative power of binary codes and efficiently improve the ANN searching accuracy.Secondly,motivated by the method of optimizing of existing hashing-based method,we propose a binary code learning approach via iterative distance adjustment.To dis-tinguish the inaccurate code pairs,we define Hamming approximating error.And our algorithm adjust the Hamming distance between code pairs by learning compact ad-ditional binary codes iteratively,so that improve the similarity preserving ability.As the experiments showed,the algorithm is able to correct Hamming spatial relationship between existing binary codes and efficiently improve the ANN searching accuracy.Finally,we apply hashing-based binary codes learning techniques to Bag-of-Feature image retrieval,improve the binary signature learning algorithm of Bag-of-Feature im-age retrieval.We propose two methods of learning binary codes as the signature of descriptor.One is based on the similarity of space distance,the other one is based on the semantic similarity.As the experiments showed,the algorithm is able to improve the similarity preserving ability of binary signature and improve the Bag-of-Feature image retrieval accuracy.