信号处理2026，Vol.42Issue(4)：596-612,17.DOI:10.12466/xhcl.2026.04.012
非对称图像检索研究综述
Asymmetric Image Retrieval:A Survey
摘要
Abstract
As deep neural networks continue to improve in terms of representational power,the accuracy of content-based image retrieval(CBIR)has increased significantly.However,the increasing model size and computational com-plexity have made deploying and applying traditional symmetric retrieval architectures at scale and in resource-constrained settings difficult.For reduced computational and communication overhead while preserving retrieval perfor-mance,balancing the efficiency and accuracy of asymmetric image retrieval—which employs models of different com-plexities and input resolutions at the query and gallery sides—has emerged as an important research topic.Nevertheless,mismatches in model capacity and input scale often induce shifts between the embedding spaces of different networks,thereby degrading matching accuracy and robustness.This paper presents a systematic review of the representative stud-ies in asymmetric image retrieval aimed at addressing these challenges and categorizes existing methods into knowledge-distillation-based and non-knowledge-distillation-based ones.For knowledge-distillation-based methods,we analyze pre-vious studies from two perspectives:single-gallery embedding space distillation and fusion embedding space distillation.The former is focused on designing distillation strategies to improve embedding alignment between query and gallery networks,while the latter is focused on constructing high-quality gallery embeddings by multi-source embedding space fusion.For non-knowledge-distillation approaches,we focus on the design principles and engineering characteristics of backward-compatible networks,neural architecture search,and network pruning for modeling cross-network feature compatibility.Finally,this paper discusses possible future research directions,including multi-scale embedding compat-ibility,structured pruning for asymmetric retrieval networks,embedding alignment between quantized and non-quantized models under edge-cloud collaboration,and adaptive retrieval strategies for dynamic scenarios,to provide guidance for future research and practical system design.关键词
深度学习/知识蒸馏/非对称图像检索Key words
deep learning/knowledge distillation/asymmetric image retrieval分类
信息技术与安全科学引用本文复制引用
谢懿,王子文,朱建清..非对称图像检索研究综述[J].信号处理,2026,42(4):596-612,17.基金项目
厦门市"揭榜挂帅"重大技术攻关项目(3502Z20251011) （3502Z20251011）
福建省科技兴警研究计划(2024Y0064) （2024Y0064）
泉州市高层次人才创新创业项目(2023C013R)Xiamen City'Jie Bang Gua Shuai'Key Technological Breakthrough Projects under Grant(3502Z20251011) （2023C013R）
Fujian Province Science and Technology Empowering Police Research Initiative(2024Y0064) （2024Y0064）
High-level Talent Innovation and Entrepreneurship Project of Quanzhou City(2023C013R) （2023C013R）
