Online Collective Matrix Factorization Hashing for Large-Scale Cross-Media Retrieval

2020 SIGIR

Di Wang XiDian University

Summary

• By developing an efficient online optimization method , take CMFH Expand to online learning mode . It incrementally updates the hash function to adapt to the changes of multimodal data flow , At the same time, generate hash code for the currently arrived data .
• Proposed OCMFH You can do this without accessing the original old data , Dynamically update the hash code of old data with the change of hash model . This can well match the hash code of new and old data , Improve retrieval performance .
• Aiming at the problem of mean change in the process of online hash learning , A zero mean strategy is proposed .

CMFH Principle

• It's batch based
• Objective function ：
  V Is the latent semantic space representation , The first is to embed features into the latent semantic space , The second is the learning of hash function .
  

OCMFH Principle

• Suppose the training data is zero Center , Then the hash function can be rewritten as
  
• zero - Mean normalization
  The original data is ：
  
  When new data blocks arrive , The average value of the whole data is ：
  
  So the new data is expressed as ：
  
  It is still used in the later formula $X{m}^{(t)}$ Instead of $\overline{X}{m}^{(t)}$
• Model update
  In the case of two modes, the formula is ：
  
  Change it to online form ：
  
• Hash code update
  The goal of matrix decomposition is to find two matrices whose products are as close to the original matrix as possible . Therefore, the product of the base matrix and the unified representation should be as close to the original data matrix as possible ：
  
  
  Therefore, the objective function of the unified representation of old data is ：
  
  Thus, an updated unified representation can be obtained without contacting the old data V, The hash code of the entire data set is updated ：
  

Experimental verification

• Data sets ：
  MIRFliker：
  image：150-dimensional edge histogram feature
  text：500-dimensional feature by performing PCA on the index vector
  NUS-WIDE：
  image：128-dimensional feature by performing PCA on its 4,096-dimensional deep feature extracted by the Caffe　implementation of　VGG Net
  text：1,000-dimensional　bag-of-words feature
  MSCOCO：
  image： 512-dimensional feature by performing PCA on its　4,096-dimensional deep feature extracted by the Caffe　implementation of VGG Net
  text：512-dimensional　feature by performing PCA on its index vector

………