An Unsupervised Face Clustering Model by Self-Enhanced Side Information

Abstract

We propose a new unsupervised face clustering scheme, in which face images are subsampled and four subimages of a quarter size of the original image are used as positive side-information to improve clustering performance. By the subsampling procedure, the number of features relevant to the classification is increased and the irrelevant features are suppressed. Consequently, the accuracy of the classification is improved. While in classical techniques the best accurate clustering rate and incorrect clustering rate stand in a small feature space, we extend this space to a larger scale of surface where the spread of data can be proper. The proposed model has led us to reduce the weight of the irrelevant components in the projection space and improve the accurate clustering performance up to 12.7%.