Automatic Face Recognition and Clustering

J. Gao

Synopsis

One of the challenges in signal processing (video or image processing) is to extract significant information from complex high-dimensional datasets, such as images. The difficulty lies in developing techniques that reduce the data in the high-dimensional space to a low dimensional approximation.

For instance the number of pixels in an image can be large, however the features in the image can be described with a few parameters and have a low-dimensional structure.  Many methods have been developed to find a low-dimensional representation of a high-dimensional data set.

Prof. Junbin Gao and his colleagues have developed a new technique which effectively organises similar data into groups automatically without human intervention. The technique was used on datasets in an unsupervised learning setting. The datasets consist of two face datasets (ORL1 and CMU PIE2) and objects datasets (COIL203).

The CMU PIE dataset, for instance, contains images of 68 subjects in different poses expressions taken under 43 different lighting conditions.

An example: images of the same person from CMU-PIE database

Applications include face recognition, image content summarization, video content structuring,

Selected Publication

Yin M, Gao J, Lin Z, Shi Q, Guo Y, "Dual Graph Regularized Latent Low-Rank Representation for Subspace Clustering", IEEE Transaction of Image Processing 2015 Dec;24(12):4918-33.