基于CNN和视觉Transformer的哈希图像检索算法综述

Survey on Hash Image Retrieval Algorithms Based on CNN and Vision Transformer

摘要: 图像检索的核心目标是从预设的图像数据库中，精准定位并提取出与给定查询图像属于同一类别的所有相关图像。然而，由于传统算法通常采用简单的线性变换来构建哈希函数，并且参数优化中需要人为手动操作。因此，传统的检索方法往往存在着较大的提升空间。近年来，深度学习和哈希技术融合在拥有高检索效率同时拥有较高的检索准确度为图像检索领域提供了新思路。本文综述了各种深度哈希方法，评估了不同类别方法的原理及特性进行介绍，对各种方法的优缺点进行分析，实验结果表明，基于深度学习的哈希图像检索方法取得了较高的检索准确性。最后展望了深度学习在优化算法和计算能力方面的潜力，预测其将在图像检索中起到越来越关键的作用，为实际应用提供更精准的技术支持。

Abstract: The core objective of image retrieval is to precisely locate and extract all relevant images belonging to the same category as a given query image from a predefined image database. However, traditional algorithms typically employ simple linear transformations to construct hash functions, requiring manual parameter optimization. Consequently, conventional retrieval methods often exhibit significant room for improvement. In recent years, the integration of deep learning and hashing techniques has provided new insights for image retrieval, offering both high retrieval efficiency and accuracy. This paper reviews various deep hashing methods, evaluates the principles and characteristics of different categories of approaches, analyzes the advantages and disadvantages of each method, and presents experimental results demonstrating that deep learning-based hashing image retrieval methods achieve high retrieval accuracy. Finally, it explores the potential of deep learning in optimizing algorithms and computational capabilities, predicting that it will play an increasingly critical role in image retrieval, providing more precise technical support for practical applications.