文献类型：期刊文献

英文题名：Deep metric learning with fine-grained features mining and orthogonalization constraint

作者：Xiao, Ting[1,2];Xu, Lingyi[1,2];Yu, Wanqian[1,2];Wang, Zhe[1,2]

机构：[1]East China Univ Sci & Technol, Key Lab Smart Mfg Energy Chem Proc, Shanghai 200237, Peoples R China;[2]East China Univ Sci & Technol, Dept Comp Sci & Engn, Shanghai 200237, Peoples R China

年份：2026

卷号：68

期号：1

外文期刊名：KNOWLEDGE AND INFORMATION SYSTEMS

收录：;EI(收录号：20261420403033);WOS:【SCI-EXPANDED(收录号:WOS:001719525100002)】；

基金：This work is supported by the Natural Science Foundation of China under Grant No.62476087, No. 62201341, and No.62306115.

语种：英文

外文关键词：Deep metric learning; Fine-grained feature mining; Semantic enhancement; Orthogonalization constraint

摘要：Deep metric learning methods typically design loss functions to learn feature embedding spaces where similar samples get closer and dissimilar samples push farther. The existing multi-proxies losses assign multiple proxies and calculate the weighted similarity between samples and proxies of the same class, which ignores the unique fine-grained representation of each sample. Moreover, simply constraining the relationship between samples and proxies may result in proxies inadvertently converging and blurring the category boundaries throughout the optimization process. This paper proposes a deep metric learning framework that integrates a fine-grained features mining module (FFMM) and an orthogonalization constraint module (OCM). The FFMM combines semantic enhancement to generate multiple proxies for each class, capturing meaningful data. For each sample, it adaptively connects to partial proxies to capture fine-grained feature representations. The OCM enforces orthogonal constraints on the relationships between non-similar proxies. These constraints guide samples to gradually move toward positive proxies while moving away from negative ones. The proposed method aims to guide appropriate movement directions for each sample by generating multiple proxies and capturing fine-grained feature representations. Furthermore, it offers a global constraint on sample movement through proxy orthogonalization. Experiments conducted on three benchmark datasets demonstrate the effectiveness of the proposed method.