Channel order estimation for multichannel blind image recovery

Several algorithms have been recently proposed for multichannel blind deconvolution of images. Many of these algorithms are critically dependent on accurate estimates of the sizes of the unknown filters - wrong estimates can result in catastrophic failure. In this paper, we present a residual-based technique for order estimation that is based on the Generalized Likelihood Ratio Test (GLRT). The performance of the algorithm is shown to be far superior to subspace rank-based methods typically used for the purpose. We conjecture that the algorithm fails only when the data is corrupted to an extent that makes any kind of reconstruction infeasible.