FLLTPlus
Some enhancement on current Fast Level Line Transform (FLLT) algorithm.
FLLT
FLLT was originally proposed by Pascal Monasse and Frédéric Guichard in 2000, by combining lower level tree
as well as upper level tree
to form a non-reduant tree of shapes (TOS)
structure of original image, with following features:
- Adaptive to scale transform, which was inevitable in segmentation etc. applications.
- Invariant to contrast change, where image's topological structure can be preserved before and after monotonous functions' acting on TOS.
Despite FLLT's brilliant performance, there still exists some limitations:
- FLLT makes no distinction on target and bakcground areas, which is controversal to our purpose, only details from human's
region of interest (ROI)
areas are needed, those non-dominant fragmentized regions could break human's visual perception process, which should be avoided. - FLLT tends to provide an
over-segmentation
image, resulting in many fragmentized shapes, especially those laid in image's background, which can hardly be noticed, but add burden on network transmission as well as local arcievement.
Procedure
Some enhancement was made in this work to overcome above mentioned issues respectively:
- Add saliency detection model with imitation of human's fixation and saccade process and extract regions with salient geometric and texture structure as target (ROI), while remaining as backgroud. Especially a MB+ saliency detection model was adopted with consideration of performance and efficiency.
- Convert image from RGB Color Space to CIE Lab Color Space and perform FLLT on image's L band.
- Merge non-dominant regions in image's background regions.
Experiment
Result of the algorithm on synthetic image can be seen as follows, with merging conditions from loose to strict.
Part of algorithm's performance on MIT300 saliency detection dataset, ranging from people, animals, architecture, artificial objects to natural scene, with merging conditions from loose to strict.
- first column is original image.
- second column is MB+ saliency detection result on original image.
- columns from three to five corresponding to image's simplification results, with merging conditions from loose to strict.
Miscellaneous
- Copyright: GNU LGPLv3
- Special thanks to reference code MegaWave2 and Pascal Monasse for the discussion.
- Contact info: hao (dot) omni (at) gmail (dot) com