Hash learning method, unsupervised online Hash learning method and application thereof

Abstract

The invention relates to a hash learning method. The method comprises steps of training the evolutionary tree through data points in the data set; obtaining a trained evolution tree; carrying out initialized Hamming code encoding on all nodes except the root node in the trained evolutionary tree, optimizing a similarity-preserving loss function of the whole evolutionary tree by using a greedy pathencoding strategy, and taking the Hamming code encoding corresponding to the minimum value of the similarity-preserving loss function as a hash code of each leaf node of the evolutionary tree; calculating an optimal matching point of a certain data point in the evolutionary tree, finding out a splitting path of a leaf node corresponding to the optimal matching point of the data point split from the root node, orderly combining hash codes of the corresponding leaf node in the splitting path of the optimal matching point of the data point, and taking the hash codes as hash codes of the data point. The invention further discloses an unsupervised online hash learning method and application. The hash method can reduce the coding complexity and has good query performance.

Description

technical field [0001] The invention relates to the field of data processing, in particular to a hash learning method, an unsupervised online hash learning method and applications thereof. Background technique [0002] With the rapid development of the Internet and various electronic devices, various types of data, such as text, images and videos, are increasing rapidly. In many application scenarios, people need to retrieve relevant content from such large-scale data. However, in large-scale data, the computation time spent to find the exact nearest neighbor for a given query point is unacceptable. In order to solve this problem, a large number of recent studies have been devoted to similar nearest neighbor (Approximate Nearest Neighbor, ANN) search. In large-scale data, the effect of ANN search can replace the exact nearest neighbor search, and the speed is very fast. ANN retrieval based on hash learning is one of the more well-known ANN retrieval technologies. It combin...

AI Technical Summary

Problems solved by technology

Supervised hashing algorithms mainly include RBM, BRE, MFH, IMH, and MLH. Although supervised hashing shows higher search accuracy than unsupervised hashing, their training requires label information. In the era of massive data , the data scale is large, the update speed is fast, and the acquisition of data labels often requires huge labor costs, so unsupervised hashing is more meaningful in practical applications

However, most unsupervised hashing algorithms need to load all the data at one time, which will take up a lot of memory, cannot be applied to streaming data, and there are few related studies

Images