A study on learning multimode image patterns

Abstract

Manifold assumption is that two nearby data points in the high-density region of a low-dimensional data manifold have the same cluster label. Manifold learning based image clustering models are usually employed at local level to deal with nonlinear manifold with an underlying assumption that better the well separated images at local level, the better will be the clustering results. Recently, it has been observed that manifold assumption might not always hold on high-dimensional image data and various clustering approaches were proposed that incorporated both local and global information to learn nonlinear manifold in image dataset. Multimode patterns in image data matrices can vary from nominal to significant due to images with different expressions, pose, illumination, or occlusion variations. Our study on learning image pattern using local neighborhood information reveals that clustering result of image clustering model varies accordingly with the distribution of image data rather than improving local neighborhood structure. Our simulation results showed that with equal well separated images at local level, performance of state-of-the-art clustering approaches is optimal for unimodal image datasets and it degrades for image data with multimodal distribution of images. We conclude that all these clustering models are based on second-order statistics and multimode patterns in image data matrices cannot be handled even by exploiting manifold information at local level.