An application of Independent Component Analysis To a Sequence of Images Data

摘要

Independent component analysis (ICA) is a statistical technique which attempts to minimize the redundancies in a set of observed data by exploiting higher than second order statistical properties. Recently, the ICA problem has become an important research and application topic in unsupervised neural learning. This technique has shown great promise in feature extraction. Here discriminatory information compression is achieved by minimizing the mutual information between the pattern vectors through a feed forward neural network and the Hebbian learning method. For a given set of pattern vectors, the mutual information between these vectors is minimized. This implies that the redundancies between original patterns are removed and the approximation patterns are (nearly) independent. The discriminatory information, which is our main interest, is compressed into weights of a set of basis vectors which span the given data. In this paper we will adapt the ICA algorithm, which was proposed by Anthony J. Bell and Terrence J. Sejnowski to determine a set of basis vectors and their corresponding weights, which span the data set of time sequence of images. First, the algorithm and its adaptation rules will be introduced. Subsequently, the ICA will be applied to extract features for the data set of a time sequence of images. Finally, performance of the ICA method will be compared to the performance of traditional feature extraction methods such as the Karhunen-Loeve expansion method.