Purpose:
To investigate the application of super resolution algorithmson images of the retina.
Methods:
With an experimental compact scanning laser ophthalmoscope (cSLO),multiple images of the retina were acquired. The images werenormalized for illumination and contrast. Rigid registrationwas performed based on automatically extracted blood vessels.Image averaging works by averaging the registered image series.Due to the random nature of noise, its amplitude is reduced.Super resolution aims at increasing the resolution of aliasedimages, in which energy in frequencies beyond the Nyquist frequencyare wrapped to lower frequencies. By analyzing multiple shiftedimages, these frequencies may be recovered.
Results:
A part of a cSLO scan is shown in figure 1, both in its originalform and after normalization and correction for anisoptropicscanning. 35 scans were averaged or combined in one super resolutionimage, with a 2 times increased resolution. Both results areshown in figure 2; the averaged image was interpolated to obtaina comparable resolution.
Conclusions:
Both the averaged and the super resolution image show detailsthat are barely visible in the individual scans. While the superresolution image appears to be sharper, it does not readilyshow smaller details. Due to scanner non-linearities, the rigidregistration fails to consistently achieve sub-pixel accuracy,which is required for super resolution. Analysis of images acquiredby a future cSLO version may show the advantage of super resolutionfor retinal imaging.Figure1. One raw scan (left) and after normalization and correctionfor anisotropy (right).Figure2. Average (left) and super resolution (right).
Keywords: image processing • retina • diabetic retinopathy