Image Clustering Using Discriminant Image Features

Abstract

Manifold learning based image clustering models are usually employed at local level to deal with images sampled from nonlinear manifold. Usually, gray level image features are used that are obtained by resizing original images through linear interpolation approach. However, significant image variance information is lost in gray level image features. Clustering models that are based on discriminant analysis can be made more effective in principal component analysis (PCA) space whereas leading projection vectors contain significant image variance information. For optimal clustering performance, we used two-dimensional two-directional PCA technique to extract significant image features. We report clustering performance of Spectral Embedded Clustering (SEC) model using discriminant image features on 6 benchmark image databases. Clustering performance is compared with existing state-of-art clustering approaches. Significant overall performance improvement is observed using proposed discriminant image features over gray level image features.