Fully Unsupervised Clustering Using Center-Surround Receptive Fields with Applications to Color-Segmentation

In this paper we argue that the emphasis on similarity-matching within the context of Content-based Image Retrieval (CBIR) highlights the need for improved and reliable clustering-algorithms. We propose a fully unsupervised clustering algorithm that is obtained by changing the non-parametric density estimation problem in two ways. Firstly, we use cross-validation to select the appropriate width of the convolution-kernel. Secondly, using kernels with a positive centre and a negative surround (DOGS) allows for a better discrimination between clusters and frees us from having to choose an arbitrary cut-off thresh- old. No assumption about the underlying data-distribution is necessary and the algorithm can be applied in spaces of arbitrary dimension. As an illustration we have applied the algorithm to colour-segmentation problems.