Picture blind source separation by auto-encoder identity mapping with structural pruning

A non-information-theoretic approach applied here for BSS of image data (pictures) is based on an auto-encoder neural network that incorporates a pruning algorithm. Nonlinear hidden units that survive the pruning will be the source extractors. The BSS state is attained as a local minimum of the error associated with the identity mapping by the auto-encoder. An internal mixing model is automatically induced in the decoder part. The BSS performance is shown to be satisfactory, including trouble cases involving noise or blanks in the mixed pictures.