A novel video coding scheme based on principal component analysis

This paper presents a novel video coding technique where most frames are represented as their projection onto a proper basis (eigenspace) computed using Principal Component Analysis (PCA). Since a video sequence contains regions with high time variations, a learning procedure is used to obtain an adequate basis. We also introduce the idea of bidirectional predicted frames to denote those frames that can be estimated from the nearest past and future PCA coefficients. Experimental results show the high quality/compression achieved using the new scheme with different eigenspace updating algorithms.