Proceedings of the 11th ACM symposium on Document engineering
| Hi-index | 0.00 |
In image clustering, digital images can be represented with a large number of visual features corresponding to a high dimensional data space. Traditional clustering algorithms have difficulty in processing image dataset because of the curse of dimensionality. Moreover, similarity between images is measured by the values of partial features. To discover clusters existing in different subspace is known as the projective clustering problem. In this paper, we propose a novel projective clustering algorithm that utilizes dense area detection in variable bin width histograms to form the description of potential cluster candidates. Those candidates with sufficient number of data objects are treated as description of clusters. Relative entropy is used as a density threshold in order to iteratively detect dense areas in each histogram. The construction of variable bin width histogram is automatic. Compared with fixed bin width histogram used in previous projective clustering algorithms, such as EPCH (an Efficient Projective Clustering technique by Histogram construction), variable bin width histogram keeps a nice tradeoff between accurately approximating the underlying distribution and clustering efficiency. Fewer input parameters are required in our proposed algorithm, and the only input parameter required is more robust to variations of other factors such as bin width and is more interpretable to general users. Experiments on an image segmentation dataset show that our algorithm has a better clustering quality than EPCH according to V-Measure.