Image Processing Reference

In-Depth Information

the accurately extrapolated frequency band to a rather small domain outside

the known data window. While we may argue that further increasing the rate

of sampling might balance the lack of accuracy, this is not the case; the only

way to extend the extrapolated region is to improve the data quality.

It was suggested and later verified that this instability can be interpreted as

being linked to the formation of super-oscillations. In others words, numerical

super-resolution algorithms construct super-oscillating signals, which opti-

mally resemble the object function. We note that this relationship between

super-resolution algorithms and optical super-resolution phenomena was

exploited from a practical point of view by constructing super-resolving filters

with the PDFT algorithm.

The strong dependence of the achievable bandwidth extrapolation on the

SNR of the data also points to a trade-off in the Lukosz sense. The space-

bandwidth product of the measured signal can be calculated straightfor-

wardly as
S
=
B
max
D
and we attempt to extract information about more than
S

independent signal features by trading for SNR. This is the same relationship

we found for super-resolving filters and super-oscillations. The exponential

growth of the required SNR as a function of bandwidth can be identified as

the main constraint for super-resolution imaging.

The interpretation of the PDFT algorithm as a “Lukosz trade-off” leaves

a pessimistic outlook on the prospect of calculating super-resolved images.

However, it has been observed that many numerical super-resolution algo-

rithms report significant improvements over the classical diffraction limit.

They account for this by distinguishing between what they call primary and

secondary super-resolution. We interpret primary super-resolution as the gain

in resolution due to exploiting the Lukosz trade-off between signal-to-noise

ratio (SNR) and image resolution. This gain in resolution is essentially inde-

pendent of the object signal and the number of samples and adheres to the

properties of the PDFT algorithm discussed so far. Secondary super-resolu-

tion is defined as any additional gain in image resolution not related to pri-

mary super-resolution. This gain may be related to other Lukosz trade-offs,

particularly in the context of image reconstruction from multiple encoded

image frames. For instance, imaging through atmospheric turbulence and

image reconstruction from multiple aliased images may be interpreted as

extracting the desired information simply from multiple separate measure-

ments. However, in each frame the different frequency bands overlap and the

information gain is transmitted similar to the degrees of freedom accessible

through primary super-resolution. It is not surprising therefore that similar

algorithms are used to recover the image and that in these cases the algorithm

displays superior image resolution simply based on the better SNR of the high-

frequency components.

However, a second source of secondary super-resolution is due to the use

of prior information, and the PDFT algorithm can be used to understand this

type of secondary super-resolution more intuitively as an interrelationship

between the estimator and the object function. In particular, we can charac-

terize the estimator in terms of the prior knowledge we inject into the recon-

struction process. Since we strive to estimate a continuous signal from a

finite set of data, we try to solve an ill-posed problem and always must make

assumptions to select the true object function from the infinite set of possible

signal reconstructions. We interpret secondary super-resolution as the result

Search WWH ::

Custom Search