New method for obtaining proper initial clusters to perform FCM algorithm for colour image clustering

• Published in: Journal of the Textile Institute 2009-04, Vol.100 (3), p.237-244
• Main Authors: Kandi, S. Gorji, Tehran, M. Amani, Rahmati, M.
• Format: Article
• Language: English
• Subjects: colour clustering; fuzzy c-mean clustering; initial centres; principal component analysis; printing pattern
• ISSN: 0040-5000; 1754-2340
• DOI: 10.1080/00405000701757545

To print a colour image on a fabric, colour image clustering is an important step to reduce the number of colours and separate the coloured pattern. Therefore, the performance of colour-clustering algorithm can strongly affect the quality of printing process. Fuzzy c-mean (FCM) clustering is a known clustering algorithm for colour image quantization but the results of FCM depend on the choice of initialization. Therefore, the problem is to find the appropriate initial centres, which are commonly chosen randomly. This paper introduces a novel initialization method for FCM algorithm for clustering colour images. We use the probability density function (pdf) of the colour image to estimate the initial centres. Considering that the three-dimensional computations are complicated and time consuming, we apply principal component analysis (PCA) to find an appropriate direction. Firstly, the three-dimensional colour points should be mapped on the first PCA vector of the colour image data. Then, to obtain the dominant colours, the pdf of the new data is calculated, the points with the highest pdf values have more chance to be initial centres. By selecting a centre, the neighbourhood data in a diameter of σ are eliminated. The number of clusters estimates the value of σ. The process is continued until approaching the desired number of colours. The experimental results show that the proposed method performs well for clustering colour images.