摘要
本文研究了一种新型的基于知识迁移的极大熵聚类技术。拟解决两大挑战性问题:1)如何从源域中选择合适的知识对目标域进行迁移学习以最终强化目标域的聚类性能;2)若存在源域聚类数与目标域聚类数不一致的情况时,该如何进行迁移聚类。为此提出一种全新的迁移聚类机制,即基于聚类中心的中心匹配迁移机制。进一步将该机制与经典极大熵聚类算法相融合提出了基于知识迁移的极大熵聚类算法(KT-MEC)。实验表明,在不同迁移场景下的纹理图像分割应用中,KT-MEC算法较很多现有聚类算法具有更高的精确度和抗噪性。
In this paper, we propose a novel technique for maximum entropy clustering （MEC） based on knowledge transfer. More specifically, we aim to solve the following two challenging questions. First, how can knowledge be appropriately selected from a source domain to enhance clustering performance in the target domain via transfer learning？ Second, how best do we conduct transfer clustering if the number of clusters in the source domain and the target domain are inconsistent.9 To address these questions, we designed a new transfer clustering mechanism called the central matching transfer mechanism, which we based on clustering centers. Further, we developed a knowl- edge-transfer-based maximum entropy clustering （KT-MEC） algorithm by incorporating our mechanism into the classic MEC approach. Our experimental results reveal that our proposed KT-MEC algorithm achieves a higher level of accuracy and better noise immunity than many existing methods when applied to texture image segmentation in different transfer scenarios.
出处
《智能系统学报》
CSCD
北大核心
2017年第2期179-187,共9页
CAAI Transactions on Intelligent Systems
基金
国家自然科学基金项目(61572236)
江苏省自然科学基金项目(BK20160187)
江苏省产学研前瞻性联合研究项目(BY2013015-02)
关键词
迁移学习
中心迁移匹配
极大熵聚类
纹理图像分割
抗噪性
transfer learning
center transfer matching
maximum entropy clustering
texture image segmentation
robustness
