Research On The Techniques Of Feature Learning And Indexing In Image Retrieval

Image retrieval is the research of finding target images from a database exactly and quickly.With the development of Internet and the spread of mobile devices,the amount of data is now growing explosively.Improving search accuracy and speed has been a great challenge for image retrieval in the era of big data when searching is performed on massive data.Focusing on image retrieval,this dissertation engages in the techniques of feature learning and indexing.The main contributions of this dissertation can be summa-rized as follows.· We propose a subspace learning method that enhances neighbor reversibility(NR)for image retrieval.Neighbor reversibility is the definition that each of two images is among the nearest neighbors of the other one.Usually,the two images that have NR correlation are strongly relevant.By enhancing the NR correlation of relevant images in subspace learning,the distribution of image features can be best refined.Experimental results demonstrate that the proposed method is effective on improving search accuracy of image retrieval.· We propose an inverted indexing method for the convolutional neural network(C-NN)features,so as to improve search speed of large-scale image retrieval when uti-lizing CNN features.The proposed method combines multiple strategies to reform inverted table for CNN features,and replaces embedding codes with hash codes to further improve search accuracy and speed.Experimental results demonstrate the effectiveness of the proposed indexing method for large-scale image retrieval.· We propose a modality-invariant image-text common feature learning method to address text-to-image retrieval.The proposed method combines two embedding networks and a modality classification network with a gradient reversal layer,and it optimizes the triplet loss and the adversarial loss simultaneously in the training phase.In this manner,the distribution difference between image embeddings and text embeddings in the common space can be best reduced.Experimental results demonstrate the effectiveness of the proposed method for text-to-image retrieval and image-to-text retrieval.· We proposes an indexing method named as multi-view cross-media hashing with semantic consistency(MCMHSC)to handle both learning and indexing of common features.The core idea of MCMHSC is to treat category as an independent view and introduces the correlation between data and categories into learning of hash functions.In this manner,category consistency of hash codes can be improved.Experimental results demonstrate that MCMHSC is superior to the existing meth-ods in terms of search accuracy and time complexity.