and has been successfully applied to restoring
parts of the Austrian film 'Opernball'. The split-screen image shows the
heavily distorted original on the left half and the restored frame on the
right.
The restoration process starts with the conversion of each frame to a digital
image with a high resolution film scanner. The size of a digital color
image is up to 45 MBytes per frame.
The next phase includes
the detection of cuts and scratches parallel to the direction of
film transport, which are the first defects to be removed. In each shot
or scene a reference frame is automatically chosen for the
brightness correction.
Furthermore a motion analysis ( local and dominant ) is performed to enable the
detection of single frame anomalies by motion compensated frame differencing
as well as camera movement necessary for detecting image vibrations. The final
restoration phase consists of the removal of single frame artefacts,
of noise suppression and image stabilisation.
While all these steps are performed
automatically, there is also the possibility for an operator to
make changes if necessary. The parameter settings
and the choice of reference frames can be modified on operator request.
The last step can be either the transformation of the image sequences
back to celluloid film or the storage on proper digital media.
Removable Classes of Artefacts:
- Dust and mud
- Image vibrations
- Mould
- Scratches
- Flicker
- Missing frames
- Color variations
- Noise
A Sample Dust Mask
The upper left image shows the original frame - maybe the artefacts due to
dust are hard to see on your display. During the restoration process the
pixel differences to a previously generated reference frame are calculated,
resulting in the dust mask (middle right image).
The pixel differences due to the actor's motion must not be taken into
account. This is achieved by tracking the motion of neighbouring pixels
through subsequent frames and compensating for pixel changes due to
the detected motion.
The pixels indicated by the dust mask are eventually restored using an
average value from the previous and next frames. The result is shown in the
bottom image.
