Completely unsupervised image segmentation using wavelet analysis and Gustafson-Kessel clustering

Image segmentation is the first step towards image analysis and image understanding. However, most image segmentation algorithms require a priori knowledge of the number of partitions in the image to be segmented. This paper introduces a novel method for completely unsupervised image segmentation by using wavelet analysis and fuzzy Gustafson-Kessel (GK) algorithm. The proposed algorithm needs no predefined number of partitions nor the number of textures in the image. The algorithm consists of feature extraction employs wavelet transform to decompose the image into different spectral components and build a feature vector for every pixel. These vectors are grouped together into clusters using the GK clustering algorithm. GK is less sensitive to fall into local minima and it has the power to generate clusters with different geometrical shapes. The appropriate number of clusters, hence number of image segments, is determined to minimize the compactness and separation clustering validity measure. The algorithm is applied to segment artificial and real images where experimental results demonstrate the effectiveness of the proposed method.