Super-resolution is an image stacking operation that uses several low resolution images to create an enhanced, higher resolution image. It seeks to utilize information embedded in
neighboring frames in a sequence to improve the resolution of a given frame. The engine does this by looking for slight
motion, and compensating for this to reveal information lost during the sampling
process.
Super-resolution works best when aliasing is taking place and the sequence involves slight motion, perhaps over a few pixels (see the image on the left below).
Aliasing is necessary because the information gained in the SR image (high-frequency content) was embedded in the low resolution images as aliasing.
Images intended for use with this engine should, therefore, not be smoothened in any way.
Animation of Frames of Original
2x Multi-frame Super-resolution
NB: The original sequence was a simulation of a shaking
camera, synthesized by shrinking a single image with slightly
different offsets.
As you can see, the engine is able to recover extra detail from the images to create a larger high-resolution image.
The “how” bit
You need at least N2 same-size images of the same scene to run this engine, where N is the scaling factor. Since the minimum scaling factor is 2, the absolute minimum number of images is 4. The images need to be aligned; this can be done by either making sure the camera remains perfectly still or by using the Align Images for Stacking feature.
This is the GUI for the Super-resolution feature:
The precision of alignment depends on how good the chosen reference point is, and in rare circumstances Chasys Draw IES will choose a bad point leading to poor alignment. If this happens, just click “Register” to choose a new point and re-align the images afresh.
The GUI offers several algorithms to choose from, as described below:
Algorithm
Details
Blending, Median
Derives destination pixels using the statistical medians of offset pixels. This method degrades very gracefully and is tolerant to low image counts, so it can be safely used with fewer than N2 images, albeit at the cost of quality. It produces images that are a little soft, so some sharpening might be required afterwards.
Blending, Average
Derives destination pixels using the statistical averages of offset pixels. This method degrades very gracefully and is tolerant to low image counts, so it can be safely used with fewer than N2 images, albeit at the cost of quality. It produces images that are a little soft, so some sharpening might be required afterwards.
Discrete Pixel Spattering
Unique to Chasys Draw IES, this technique gives the sharpest, best quality results, but only if the data available in the input images is sufficient – if not, it tends to produce artifacts. Destination pixels are selected using a proprietary voting algorithm. It is recommended that you use more than 2×N2 images when using this option; for example, for 3×3 scaling, use 2×(32) = 2×(9) = 18 images or more.
To
maximize on the information available to the super-resolution engine, the input images must not be smoothened in any way. If using video, it should be recorded uncompressed.
