Filtering of map images by context tree modeling Pavel Kopylov and Pasi Fränti UNIVERSITY OF JOENSUU DEPARTMENT OF COMPUTER SCIENCE FINLAND
Noisy map images Original image (4 colors) Distorted image (1931 colors) Noise can originate from scanning, changing resolution, lossy JPEG compression.
Context-based filter Estimate pixel probability relatively to context Neighborhood configuration defined by a local template
Sample statistics (part 1)
Sample statistics (part 2)
Example
Context tree
Context tree construction
Test material
Experiments Apply impulsive or content-dependent noise to the original image. Apply filtering. Compare performance: Euclidean distance between two color samples in uniform L*a*b* (CIELAB) space
Impulsive noise
Content-dependent noise
Impulsive noise OriginalNoisy Context treeVector Median
Content-dependent noise OriginalNoisy Context treeVector Median
Example OriginalVector MedianContext tree
Impulsive noise OriginalVector MedianContext tree
Content-dependent noise OriginalVector MedianContext tree
Conclusions Capable of utilizing larger neighborhood than fixed-size template. The method outperforms vector median filter when noise level <25%. Tree construction requires extensive amount of memory; future work is needed to optimize this part.