An easily computable eight times overcomplete ICA method for image data

Here we present a procedure for finding an eight times overcomplete ICA description of image data using the symmetries defined by the rigid motions of a square. The procedure for estimating the basis requires only a small change to any classic ICA procedure and the data representation in this overcomplete description is unique. Coding and decoding in this description are essentially as easy as in classic ICA methods. We also show that this. description is genuinely more sparse than a non-overcomplete ICA method.