A Cluster Center Initialization Method using Hyperspace-based Multi-level Thresholding (HMLT): Application to Color Ancient Document Image Denoising

Many iterative supervised clustering algorithms such as K-means and its derivatives depend closely on the initial cluster center positions. In order to overcome the convergence problems inherent in the clustering algorithms (i.e., local optimum), and subsequently to avoid a drop in clustering performance, many researchers continue to propose novel efficient methods able to determine automatically the optimal cluster centers. Therefore, in this paper, we propose a simple and efficient cluster center initialization method, called hyperspace-based multi-level thresholding (HMLT). The proposed HMLT method is based on using a novel multi-level thresholding approach on the multi-dimensional representation of color images (called hyperspace). In order to show the high performance of the HMLT method, experiments have been conducted using a recent clustering method, called the hyperkernel-based intuitionistic fuzzy c-means (HKIFCM), and after initializing the cluster center positions randomly and by means of the HMLT method. The HKIFCM clustering method that its performance tightly depends on the cluster center initialization, is applied for color ancient document image denoising (i.e., separate noise from text and background). Qualitative and quantitative assessments of results are deduced from a number of ancient document images collected from two different datasets.