Concluding Remarks - Sparse Representations and Compressive Sensing for Imaging and Vision - page 92

Number of training samples : The methods presented in [156] and [112]
harnessing sparsity are very effective yet they suffer from some limitations. For
instance, for good recognition performance, the training image set is required to
be extensive enough to span the conditions that might occur in the test set. For
example in the case of face biometric, to be able to handle illumination variations
in the test image, more and more training images are needed in the gallery. But
in most realistic scenarios, the gallery contains only a single or a few images of
each subject and it is not practical to assume the availability of multiple images
of the same person under different illumination conditions. Another limitation
of this approach is that the large size of the matrix, due to the inclusion of the
large number of gallery images, can tremendously increase the computational
as well as the storage complexity which can make the real-time processing very
difficult. Can sparsity motivated dictionary learning methods offer solution to
this problem?

￿