Add undecimated pyramid where there is no downsampling.
It is useful when translation invariance is wanted or if analyzing upper levels --lower resolution-- and we are interested in knowing where, in the domain of the input image, is the feature, without needing any re-scaling. Note that the amount of information between regular and undecimated pyramids is the same (Shannon theorem). But the undecimated generates all its wavelet coefficients and low-pass residual with the same size than the input image. This makes it more expensive, computationally and in memory.
Add undecimated pyramid where there is no downsampling.
It is useful when translation invariance is wanted or if analyzing upper levels --lower resolution-- and we are interested in knowing where, in the domain of the input image, is the feature, without needing any re-scaling. Note that the amount of information between regular and undecimated pyramids is the same (Shannon theorem). But the undecimated generates all its wavelet coefficients and low-pass residual with the same size than the input image. This makes it more expensive, computationally and in memory.