Research On Image Retrieval Algorithm Based On Deep Hashing

In the era of the Internet,a massive amount of image data has emerged.How to efficiently conduct similarity retrieval from these data has become a crucial issue.Nearest neighbor search has high accuracy,but it comes with high computational and storage complexity,making it difficult to meet the demands of real-time retrieval and large-scale data storage.In contrast,hash methods has significant advantages in computation and storage,so it is widely used in large-scale image retrieval.The main idea of hashing retrieval tasks is to map high-dimensional features in the original space to compact binary codes in Hamming space,minimizing the Hamming distance between similar data pairs while maximizing the Hamming distance between dissimilar data pairs.The main work of this thesis is aimed at how to generate high-quality hash codes and improve the accuracy of image retrieval as follows.1.Addressing the shortcomings of hash methods based on pairwise similarity/triplet similarity,this thesis proposes a deep hashing algorithm based on amended supervised contrastive learning.On one hand,it increases coverage of the similarity relationship in the dataset by expanding sample size;on the other hand,it amend supervised contrastive learning by introducing label set intersection and union ratio.The algorithm encourages difficult samples with higher label set intersection and union ratios with anchor images to contribute larger gradients,thus achieving accurate mining of difficult positive samples.In addition,a slice-based Wasserstein distance is used to define quantization loss in order to promote distribution of continuous output values from hash functions that can approach discrete uniform distribution.This allows low-dimensional Hamming space to be fully utilized and improves model retrieval performance.2.In response to the problem that the hash center of current center-similaritybased hashing algorithms cannot be learned,this thesis proposes a deep hashing algorithm based on label representation learning..For single-label retrieval tasks,there is no dependency relationship between different labels,and the algorithm encourages individual label representations to stay away from each other and make full use of the low-dimensional embedding space;for multi-label retrieval tasks,the algorithm takes into account the co-occurrence relationship between different labels and transforms it into the similarity relationship between label representations in the embedding space based on graph convolutional networks.In the image feature learning stage,each image is encouraged to approach its corresponding label representation in the embedding space,so that similar images can approach each other.In this thesis,two innovative algorithms are proposed for the deep hash image retrieval task.Through extensive experimental results,it is shown that the algorithms proposed in this thesis bring effective improvement in retrieval performance on both single-label image retrieval tasks and multi-label retrieval tasks.