Qin Xiaoqian, Liu Dakun, Gui Bin
School of Geography and Planning, Huaiyin Normal University, Huai'an, Jiangsu, China.
School of Mechanical Engineering UGS College, Yancheng Institute of Technology, Yancheng, Jiangsu, China.
PLoS One. 2025 Sep 9;20(9):e0329574. doi: 10.1371/journal.pone.0329574. eCollection 2025.
Previous studies have demonstrated that metric learning approaches yield remarkable performance in the field of kinship verification. Nevertheless, a prevalent limitation of most existing methods lies in their over-reliance on learning exclusively from specified types of given kin data, which frequently results in information isolation. Although generative-based metric learning methods present potential solutions to this problem, they are hindered by substantial computational costs. To address these challenges, this paper proposes a novel correlation calculation-based multi-task learning (CCMTL) method specifically designed for kinship verification. It has been observed that kin members often exhibit a high degree of similarity in key facial organs, such as eyes, mouths, and noses. Given this similarity, similar facial features between kin members with different kin relationships frequently demonstrate certain correlations. Inspired by this observation, our proposed method aims to learn a set of metrics by leveraging both the specified kinship data and the correlations among various kinship types. These correlations are determined through an in-depth investigation of the spatial distribution relationship between the specified kinship data and other kinship types. Furthermore, we develop an efficient algorithm within the multi-task learning framework that integrates correlation exploitation with metric learning. This innovative approach effectively resolves the issue of information isolation while minimizing computational overhead. Extensive experimental validation conducted on the KinFaceW dataset demonstrates that the proposed CCMTL method achieves superior or comparable results to those of existing methods.
先前的研究表明，度量学习方法在亲属关系验证领域表现出色。然而，大多数现有方法的一个普遍局限性在于它们过度依赖于仅从特定类型的给定亲属数据中学习，这经常导致信息孤立。尽管基于生成的度量学习方法为这个问题提供了潜在的解决方案，但它们受到大量计算成本的阻碍。为了应对这些挑战，本文提出了一种专门为亲属关系验证设计的基于新颖相关性计算的多任务学习（CCMTL）方法。据观察，亲属成员在关键面部器官（如眼睛、嘴巴和鼻子）上往往表现出高度相似性。鉴于这种相似性，具有不同亲属关系的亲属成员之间的相似面部特征经常表现出一定的相关性。受此观察启发，我们提出的方法旨在通过利用指定的亲属关系数据和各种亲属关系类型之间的相关性来学习一组度量。这些相关性是通过对指定亲属关系数据与其他亲属关系类型之间的空间分布关系进行深入研究来确定的。此外，我们在多任务学习框架内开发了一种高效算法，该算法将相关性利用与度量学习相结合。这种创新方法有效地解决了信息孤立问题，同时将计算开销降至最低。在KinFaceW数据集上进行的广泛实验验证表明，所提出的CCMTL方法取得了优于或与现有方法相当的结果。
