Research And Application Of Incremental Hashing Based On Sample Selection For Image Retrieval In Dynamic Environment

In current big data era,image data increases every day.Therefore,a fast image retrieval algorithm based on similarity and with a compact storage is very important.Hashing method has a high retrieval efficiency and a low computational cost.So,hashing methods have been widely applied in large scale image retrieval problems.Most existing hashing methods assume that the database for image retrieval is fixed.However,the real data environment is usually dynamic.Therefore,static hashing methods are often difficult to adapt changes in a dynamic data environment and result in continuous declines in their image retrieval performances.New images being added to the database may not be sampled from the same data distribution of images in the current database which results in a change in the data distribution of the database,i.e.concept drift.At present,online hashing methods have been proposed to adapt changes in the dynamic data environment.However,these methods only assume emerging images obey a certain fixed distribution and cannot deal with scenarios with concept drifts in the dynamic data environment.Incremental hashing(ICH)is an effective method to deal with the concept drift problem in dynamic data environment.ICH trains a new hash table using the latest data batch and weights multiple hash tables based on their performances on the latest data to dynamically update the hash tables to adapt changes in the data environment.However,ICH uses all images in the latest data batch to train a new hash table may be redundant and inefficient.Not all images in the new data batch are useful for new hash table training.Some images may even be duplicate or highly similar to existing images in the database.To relieve aforementioned problems in ICH,two incremental hashing methods based on sample selection from different perspectives are proposed.The main idea of the proposed methods is to select representative samples with database distribution information for each class from both new and old image datasets to train new hash functions.These two methods use multiple hash tables to preserve the similarity information between images at different times and adaptively update hash tables with their corresponding weights according to the latest dynamic data environment.For query images,retrieval results are rendered by a weighted reranking of image retrieval results from multiple hash tables.A variety of different dynamic data scenarios are simulated using three real image datasets.Experimental results show that these two proposed methods outperform other mainstream hashing methods inimage retrieval in dynamic data environments.