Research On Image Retrieval Based On Deep Hashing

In recent years,internet technology has brought about great changes in our lives.As an important information carrier,the number of images is also increasing exponentially.How to effectively and quickly find interesting images is a valuable research topic in the massive image data.Hashing-based image retrieval algorithms have quickly attracted many researchers' attention due to its low storage space and fast calculation speed.With the development of deep learning,lots of hashing algorithms based on deep learning have made certain achievements in the field of image retrieval,but there is still room for improvement for different tasks.In order to further improve the performance of image retrieval,this paper combines supervised and unsupervised aspects to carry out the work.The main of research contents is as follows:1.This paper proposes a two-stream supervised deep hashing with joint classification loss.The algorithm mainly includes a two-stream supervised deep framework.The first stream neural network focuses on the spatial semantic relevance learning,and the second stream neural network focuses on the channel semantic correlation learning.The two neural networks are integrated into a unified framework,and the triplet loss function,classification loss function and quantitative loss function are used to optimize them at the same time.The learning process of the image space stream and the image channel stream benefit each other,thereby generating discriminative hash codes.2.For the problem that the defects of k-nearest neighbor algorithm and data imbalance in the training process,this paper proposes an unsupervised deep hashing model based on extended k nearest neighbors.In the model,the automatic encoder is used as the training network,and uses the extended k nearest neighbor algorithm to construct similarity matrix to guide the model learning.In the training process,an adaptive weight is added to the loss function to solve the problem of data imbalance.Finally,a lot of experiments on three datasets have proved its effectiveness.3.In order to solve the problem of pseudo-labels generated by the k-means clustering algorithm in the current unsupervised deep hashing algorithms,this paper proposes an unsupervised deep hashing algorithm based on iterative optimization of pseudo-labels.During the training process,the model continuously updates the network parameters,re-clusters and optimizes the pseudo-labels,which avoids the problem of unstable pseudo-labels in early feature clustering.The network parameters updated in the previous round of training are used to optimize the pseudo-labels in the next round of feature clustering,and iterations are repeated until the optimal results are obtained.