Texture Segmentation Based on Voting of Blocks, Bayesian Flooding and Region Merging

1
Texture Segmentation Based on Voting of Blocks,
Bayesian Flooding and Region Merging
C. Panagiotakis(1), I. Grinias(2) and G.
Tziritas(3)
Presenter Dr. Costas Panagiotakis, Assistant
Professor, (1) Business Administrator
Administration Dep., TEI of Crete, Agios
Nikolaos, Greece
(2) Geoinformatics and Surveying Dep., TEI of
Serres, Serres, Greece(3) Computer Science
Department, University Of Crete, Greece
22th International Conference on Pattern
Recognition
07-07-2009
2
Introduction Related Work

• Segmentation of images is quite important for
  many applications, such as content based image
  retrieval and object recognition.
• In our previous work 1, we proposed a framework
  that performs automatic segmentation of images,
  knowing only the number of regions, which
  involves feature extraction and classification in
  feature space, followed by flooding (PMCFA) and
  merging in spatial domain.
• PMCFA has been also successfully applied on
  interactive image segmentation 2, where the
  goal is to classify the image pixels into
  foreground and background classes, when some
  foreground and background markers are given.

1 C. Panagiotakis, I. Grinias and G. Tziritas,
Natural Image Segmentation based on Tree
Equipartition, Bayesian Flooding and Region
Merging, IEEE Transactions on Image Processing,
Vol. 20, No. 8, pp. 2276 - 2287, Aug. 2011. 2
C. Panagiotakis, H. Papadakis, E. Grinias, N.
Komodakis, P. Fragopoulou and G.
Tziritas, Interactive Image Segmentation Based on
Synthetic Graph Coordinates, Pattern Recognition,
vol. 46, no. 11, pp. 2940-2952, Nov. 2013.
3
Introduction Contribution

• The proposed method uses features that are
  optimized and tested for textured images.
• We solve the problem to find subset of blocks
  that represent well the whole dataset of blocks
  by a new framework that takes into account the
  blocks similarity and topology. The
  representative blocks are used to extract the
  features for each class.
• The proposed method automatically computes the
  number of classes regions by a new criterion that
  takes into account the average likelihood per
  pixel of the classification map and penalizes the
  complexity of the regions boundaries. In 1-2
  the number of classes were given.

1 C. Panagiotakis, I. Grinias and G. Tziritas,
Natural Image Segmentation based on Tree
Equipartition, Bayesian Flooding and Region
Merging, IEEE Transactions on Image Processing,
Vol. 20, No. 8, pp. 2276 - 2287, Aug. 2011. 2
C. Panagiotakis, H. Papadakis, E. Grinias, N.
Komodakis, P. Fragopoulou and G.
Tziritas, Interactive Image Segmentation Based on
Synthetic Graph Coordinates, Pattern Recognition,
vol. 46, no. 11, pp. 2940-2952, Nov. 2013.
4
(No Transcript)
5
Methodology Feature Selection

• The image is divided into overlapping blocks
  (50 overlapping).
• 64 64 block for a frame of 512 512 pixels is
  used.
• We use the three components of Lab color space
  to represent the color
• The last component is the energy of horizontal
  and vertical components from wavelet transform
  using the fourth-order binomial filter 1 4 6 4
  1/16. We show that these components of wavelet
  transform suffice to represent well the texture
  information.

6
Methodology MAXR BLOCKS SELECTION

• Goal Select the MAXR most representative image
  blocks taking into account the blocks similarity
  and topology.
• Main Steps
• The M image blocks are represented by a graph G,
  whose weights are given by the Mallows distance
  of three color components and of the texture
  component of the corresponding blocks
  (4-connections neighborhood).
• Next, we find the MxM matrix of all shortest
  paths in graph G taking into account similarity
  and topology.
• Similar results are also obtained and by using
  the MST of G instead of G.

MxM matrix of all shortest paths in graph G
7
Methodology MAXR BLOCKS SELECTION

• The proposed MAXR BLOCKS SELECTION is inspired
  from 1.
• Main Steps
• The first block is given by the block of
  minimum mean distance from others (centroid).
• Next, we repeat MAXR-1 times the following
  procedure
• The next block is given taking into account the
  current selected blocks.
• We get the block that has low distances from
  others (non selected blocks) and high distance
  from the selected blocks.

MAXR Selected Blocks
4
3
1
6
5
2
1 C. Panagiotakis, Clustering via Voting
Maximization, Journal of Classification, 2014
(accepted).
8
Flooding Process for Class Propagation PMCFA (1/3)

• Definition of a topographic map for each class k
  using the computed conditional probabilities
• Height of pixel s represents the dissimilarity of
  s from class k, defined as
• ln Pk?(s)
• where Pk?(s) is the a-posteriori probability
  of class k given the feature vector ?(s).

Class
Class
1 C. Panagiotakis, I. Grinias and G. Tziritas,
Natural Image Segmentation based on Tree
Equipartition, Bayesian Flooding and Region
Merging, IEEE Transactions on Image Processing,
Vol. 20, No. 8, pp. 2276 - 2287, Aug. 2011.
9
Flooding Process for Class Propagation PMCFA (2/3)

• Path cost Ci(s0,s) between pixels s and s0 the
  maximum height of pixels in that path.
• Topographic distance dk(s) between s and s0 the
  minimum cost of paths between s and s0.

1 C. Panagiotakis, I. Grinias and G. Tziritas,
Natural Image Segmentation based on Tree
Equipartition, Bayesian Flooding and Region
Merging, IEEE Transactions on Image Processing,
Vol. 20, No. 8, pp. 2276 - 2287, Aug. 2011.
10
Flooding Process for Class Propagation PMCFA (3/3)

• Input
• Topographic map and
• initial regions of high confidence per class.

• Priority Multi-Class Flooding Algorithm
• Competitive growing for both the computation of
  topographic map and pixel labeling.
• Flooding stops when all image pixels are labeled.

Class
PMCFA Result
Original image
Topographic map
1 C. Panagiotakis, I. Grinias and G. Tziritas,
Natural Image Segmentation based on Tree
Equipartition, Bayesian Flooding and Region
Merging, IEEE Transactions on Image Processing,
Vol. 20, No. 8, pp. 2276 - 2287, Aug. 2011.
11
Merging Process

• Usually, the number of computed regions is
  greater than the real number of classes.
• A merging state solves this over-segmentation
  problem.
• We have used a greedy algorithm that iteratively
  merges the regions taking into account the
  dissimilarity in appearance of the segments and
  the gradient on region boundaries.

12
Selection of the appropriate segmentation Map

• We select the segmentation that minimizes a
  criterion C(k) FS(K) ? PC(k) taking into
  account
• the average likelihood per pixel of the
  classification map (FS(K)) and
• penalizes the complexity of the regions
  boundaries (PC(K)) that is computed from the
  points with curvature higher than 0.5 multiplied
  by a normalization factor.

13
Experimental Results on Prague Texture
Segmentation Benchmark
14
Conclusions

• An unsupervised segmentation algorithm is
  proposed which combines
• color and texture features,
• region features and
• topology.
• yielding high performance results.
• Results on Prague Texture Segmentation Benchmark