Learning interpretable binary codes via semantic alignment for customized image retrieval

doi:10.1016/j.patcog.2025.112380

	Learning interpretable binary codes via semantic alignment for customized image retrieval
	Qiao, Shishi 1,2,3; Wang, Ruiping 2,3; Chen, Xilin 2,3
	2026-04-01
发表期刊	PATTERN RECOGNITION
ISSN	0031-3203
卷号	172 页码:14
摘要	The single modality hashing (SMH) has achieved impressive performance on image retrieval task in recent years. The only fly in the ointment is that most of the methods mainly measure the image similarity based on the highlevel class labels. The retrieval needs in the real world are diverse in form of different subsets of the semantics (not only the category labels) presented in the query image. However, existing SMH methods fail to account for such customized image retrieval task that allows users to select visual semantics or their combinations present in the query and retrieve similar images based on such selected semantic descriptions. To address such practical issues, we propose a deep hashing to learn Interpretable Binary Codes (IBC), endowing the hashing bits with semantic interpretability rather than purely entangling the class information in the whole codes, i.e., aligning the criteria of binary space partition of each bit with a particular visual semantic concept. Specifically, binary encoding is a highly non-linear operation of dimension reduction, the semantic and spatial information of which has respectively been abstract and lost heavily. In light of the rich semantic interpretability and binary concept detection ability of convolutional filters, we innovatively transfer the semantic knowledge from filters to hashing bits by align the distributions of the binary codes and filter activations that capture the presence/absence of visual patterns in images. To further improve the semantics of filters/bits, the shared and learnable classification rules are introduced and optimized to disentangle the sparse composition between the category label and encoded semantics in filters/bits. With high interpretability, we can selectively combine bits corresponding to the target semantics during retrieval, thereby enabling flexible and customized similarity searches. Extensive experiments on several large-scale datasets covering general objects and scenes, single and multiple label scenarios, demonstrate the interpretability and functionalities of learned binary codes for the customized image retrieval tasks.
关键词	Image retrieval Hashing Interpretability Semantic alignment
DOI	10.1016/j.patcog.2025.112380
收录类别	SCI
语种	英语
资助项目	National Key R&D Program of China[2021ZD0111901] ; Natural Science Foundation of China[U21B2025] ; Natural Science Foundation of China[U19B2036] ; Natural Science Foundation of China[61922080] ; Natural Science Foundation of China[62206260] ; Natural Science Foundation of Shandong Province[ZR2024QF077]
WOS研究方向	Computer Science ; Engineering
WOS类目	Computer Science, Artificial Intelligence ; Engineering, Electrical & Electronic
WOS记录号	WOS:001566676700011
出版者	ELSEVIER SCI LTD
引用统计
文献类型	期刊论文
条目标识符	http://119.78.100.204/handle/2XEOYT63/41725
专题	中国科学院计算技术研究所期刊论文_英文
通讯作者	Wang, Ruiping
作者单位	1.Ocean Univ China, Fac Informat Sci & Engn, Qingdao 266100, Peoples R China 2.Chinese Acad Sci, Inst Comp Technol, Key Lab Intelligent Informat Proc, Beijing 100190, Peoples R China 3.Univ Chinese Acad Sci, Beijing 100049, Peoples R China
推荐引用方式 GB/T 7714	Qiao, Shishi,Wang, Ruiping,Chen, Xilin. Learning interpretable binary codes via semantic alignment for customized image retrieval[J]. PATTERN RECOGNITION,2026,172:14.
APA	Qiao, Shishi,Wang, Ruiping,&Chen, Xilin.(2026).Learning interpretable binary codes via semantic alignment for customized image retrieval.PATTERN RECOGNITION,172,14.
MLA	Qiao, Shishi,et al."Learning interpretable binary codes via semantic alignment for customized image retrieval".PATTERN RECOGNITION 172(2026):14.