| In the information age,search engines always face large-scale media data,which puts great pressure on the retrieval speed of engines.Hash methods have attracted the attention of academic scholars and industrial experts because of their fast retrieval efficiency,high storage efficiency and small storage space.Hashing methods aim to represent high-dimensional data as compact binary hash codes in Hamming space that maintain the original similarity.Based on the hash code representation,the engine is able to speed up the search by linear scan or hash table lookups.And with the rapid development of deep learning technology,deep supervised hashing has shown extremely good retrieval accuracy in the face of large-scale media retrieval.However,deep supervised hashing nowadays also faces challenging problems such as information loss in the dimensionality reduction and relaxation process,and extremely sparse data distribution in Hamming space retrieval.In addition,how to apply hashing techniques in image retrieval to cross-modal retrieval,so that different modal features are mapped to hash codes for cross-modal retrieval,which is also becoming an important research trend.This thesis focuses on the problems related to deep hashing algorithms for image retrieval and cross-modal retrieval,and the main works include the following:1.a deep hashing algorithm combining category information and cosine metric supervision is proposed for image retrieval tasks.The algorithm uses a cosine distance metric to generate real hash codes to reduce the similarity loss of mapping highdimensional features into binary hash codes,and constrains the quantization loss caused by the relaxation strategy through L2 regularity.In addition,this algorithm enables the generated hash codes to restore the category information by adding linear transformation,which reduces the difference between the similarity of hash codes and the actual semantic similarity,and the similarity preservation ability is improved,which in turn improves the retrieval accuracy.2.Most of the previous deep hashing methods focus on maximizing the retrieval performance for linear scan.However,the complexity of linear scan is still unacceptable for very large datasets.Hamming space retrieval can achieve constant time complexity by hash table lookups.However,previous deep hashing methods fail in Hamming space retrieval as the data distribution becomes sparse as the Hamming space increases and the data points hardly fall within a given small Hamming radius used for interception.To address this problem,this paper proposes an exponential hashing algorithm with a distinguishability penalty.The algorithm introduces a discriminative penalty in the exponential metric similarity loss function for similar/dissimilar data inside and outside the Hamming sphere to optimize the Hamming space,thus improving the effectiveness of Hamming space retrieval.Extensive experiments demonstrate that the algorithm achieves excellent performance in the benchmark dataset.3.Building on the two previous works,it is further extended from image retrieval to cross-modal retrieval.In order to fully utilize the category information to simultaneously maintain the ranking performance of cross-modal deep hash retrieval and the aggregation capability of similar data,this paper proposes a category information-guided graph neural network cross-modal retrieval algorithm for crossmodal retrieval.To enrich the category information of hash codes,this algorithm uses category relations for graph model construction,and uses graph convolutional networks and multimodal factorized bilinear pooling(MFB)to further integrate the category information into image networks and text networks.In addition,inspired by the above,the algorithm also uses the distinguished exponential hash as a loss function to maintain the effectiveness of Hamming space retrieval while exploiting the similarity relationship between graphs and texts for graph model construction,and uses graph attention networks to interactively fuse information on graph and textual real hash codes to obtain category representations.The algorithm achieves excellent performance through comparative experiments on publicly available datasets. |
