Clustering of different shapes of the same object has an inordinate impact on various domains, including biometrics, medical science, biomedical signal analysis, and forecasting, for the analysis of huge volume of data into different groups. In this work, we present a novel shape-based image clustering approach using time-series analysis, to guarantee the robustness over the conventional clustering techniques. To evaluate the performance of the proposed procedure, we employed a dataset consists of various real-world irregular shaped objects. The shapes of different objects are first extracted from the entire dataset based on similar pattern using mean structural similarity index. Furthermore, we performed radical scan on the extracted shapes for converting them to one-dimensional (1D) time-series data. Finally, the time series are clustered to form subgroups using hierarchical divisive clustering approach with average linkage, and Pearson as distance metrics. A comparative study with other conventional distance metrices was also conducted. The results established the superiority of using Pearson correlation measure, which provided the maximum F1-score with exact number of shapes under a sub-cluster, while the corresponding outcomes of other approaches results in a poor and inappropriate clustering.

同一对象的不同形状的聚类会对包括生物识别，医学，生物医学信号分析和预测在内的各个领域产生过分的影响，以便将大量数据分析成不同的组。在这项工作中，我们提出了一种使用时间序列分析的新颖的基于形状的图像聚类方法，以确保优于常规聚类技术。为了评估所提出程序的性能，我们采用了由各种现实世界中不规则形状的对象组成的数据集。首先使用均值结构相似性指标基于相似模式从整个数据集中提取不同对象的形状。此外，我们对提取的形状进行了彻底扫描，以将其转换为一维（1D）时间序列数据。最后，使用具有平均链接的分层分裂聚类方法将时间序列聚类为子组，并以Pearson作为距离度量。还进行了与其他常规距离度量的比较研究。结果建立了使用Pearson相关测度的优越性，该测度提供了最大F1分数，并且在子集群下具有准确的形状数量，而其他方法的相应结果导致较差且不合适的聚类。