John VanSickle wrote:
> 
> Now you could make a huge matrix (N x N, where N is the number of 
> elements in the image to be sharpened) and solve it.  But for an image 
> of, say, typical embedded YouTube size, you are looking an matrix that 
> is on the order of 80K x 80k elements.  Granted, most of the elements 
> are zero, but it's still a headache.

In the particular case of deconvolution there are much more efficient 
algorithms available.  There's various techniques, but one insight that 
is commonly exploited is that convolution in the spatial domain is 
equivalent to multiplication in the Fourier domain.  This in the absence 
of noise (which is a *big* assumption) and ignoring image boundaries you 
could compute (non-blind) deconvolution by using FFT on the image and 
the kernel and dividing, which is a pretty efficient operation.

IIRC the approach above isn't used much in practice, largely because of 
the noise assumptions, but some of the more sophisticated methods still 
perform operations in frequency space for efficiency.

Post a reply to this message