Auxiliary Materials for Best Analogs for Replacing Missing Image Data
by Chen Ning and Cliff Reiter

Best Analogs forecasting is a technique for using local patterns to extrapolate missing data. That basic idea can be applied to missing image data. Several details should be changed to obtain a quicker, better algorithm that we describe as the standard algorithm in the manuscript "Best Analogs for Replacing Forecasting Missing Image Data". As an illustration of our standard algorithm, we can apply our standard algorithm to the following image, with missing data marked in magenta:

180k

The result is:

180k

A movie showing the evolution of the standard forecast:

5.6M
If your browser does not directly support viewing this *.mov file, save it locally and view it with Quicktime™.

Some comparisons showing sensitivity may be seen as follows. Notice that while the details vary, the general form of the forecast is not dramatically different.
180k The standard forecast
186k The standard forecast changed to use 2 repetitions of the diagonal passes instead of three
186k The standard forecast changed to use all weights of one.
186k The standard forecast changed to use weights of 8 and 4 in place of 4 and 2.

A paper by JC Sprott (which may be found [here]) used a probabilistic automaton to fill in missing image data. He applied his algorithm to an illustration of a cat Table: the cat, his forecast, standard forecast, a better one.
132k
The cat with missing data
132k
The cat with missing data replaced using Sprott's probabilistic automaton
197k
The cat with missing data replaced by the standard algorithm
132k
The cat with missing data replaced using Sprott's probabilistic automaton


See also:
· A Preprint of the Best Analogs for Replacing Missing Image Data
· JC Sprott's [paper] on missing image data.
· J implementation of time series fractal forecasting [abstract].