Large-scale Image Retrieval Base On Multi-feature Fusion Hashing

In recent years,due to its advantages of small storage space and fast retrieval speed,hashing has gradually become the most widely studied retrieval algorithm in the fields of large-scale image retrieval.The essence of the hashing method is to transform the high-dimensional data in the original space into the low-dimensional binary hash codes in the Hamming space.Hash codes has the advantages of simple and convenient storage,so it can be quickly and efficiently complete retrieval and storage tasks in large-scale image retrieval.Although existing multi-feature hashing methods have made some progress,they still have some problems: 1)Multi-feature fusion hashing needs to consider the complementarity between multiple features,which play an important role in the retrieval performance of hash codes.However,traditional methods fusion multiple features with fixed weights,which requires an extra hyper-parameter to balance the regularization terms.However,when the semantic label is unknown,The hyper-parameter adjustment process will consume huge manpower.2)Semantic annotation of multi-feature image data is time-consuming and labor-intensive,especially for large data sets,it is more difficult to obtain the label information of all data in the data set.3)The optimization speed is slow.At present,several discrete multi-feature hashing methods are used to solve the hash code bit by bit through the discrete cyclic coordinate descent method,and the optimization speed is slow.In this paper,we directly learn discrete multi-feature hash codes with fast optimization speed and low storage cost through iterative optimization.In response to the above problems,this paper studies a large-scale image retrieval method based on multi-feature fusion hashing.The main contents are as follows:(1)This paper proposes a fusion weight adaptive multi-feature hash method.The model has designed a unified learning framework,which can quickly and adaptively fuse multiple image features to effectively carry out large-scale image retrieval.Specifically,the method first uses neural networks to extract the depth feature representation of the image,then designs a adaptive weighting scheme to effectively merge multiple features,and finally achieves the goal of fast and accurate large-scale image retrieval.Experiments have been conducted on the public datasets,and the experimental results prove that the method proposed in this paper has better performance than the latest related methods.(2)This paper proposes a fast semi-supervised multi-feature hashing method,which only needs to partially mark the training data to complete the retrieval task.Specifically,after obtaining a deep feature representation,this method first performs a non-linear transformation on labeled and unlabeled data to preserve the similarity between samples,and then uses asymmetric supervised learning for labeled data to learn more discriminative hash code,finally,the hash function is learned by combining labeled data and unlabeled data.Experimental results on multiple public datasets prove the superiority of the proposed method.