Extracting features based on independent component analysis with source dependency
Wei Qu, Heping Liu, Zhang Hai-Jun · 2005
This paper addresses the problem of representing the image and speech signal using a set of features that are approximately statistically independent. This statistical independence simplifies building probabilistic models based on these features that can be used in applications like speaker recognition. Since basic independent component analysis (ICA) isn't suitable to many applications because of the sources' assume that they are i.i.d., we modeled the dependency by a non-linear function, and a multi-layer feed-forward neural network was used to implement the non-linear ICA algorithm, i.e. SD-ICA, which has low computational complexity and fast convergence. The experiment given later proves that the algorithm can be used in extracting both images and speech features and it outperforms than basic ICA.