Colour image processing for SHADOW REMOVAL

1
Colour image processing for SHADOW REMOVAL

• Alina Elena Oprea, University Politehnica of
  Bucharest
• Katarzyna Balakier, Fundacion SENER
• Weronika Piatkowska, Jagiellonian University
• Alexandru Popa, Technical University of
  Cluj-Napoca  

2
Alexs angels team
Weronika Alex Alina Kasia
3
Layout

• Problem statement
• The System Overview
• Simulations and Results
• Future Perspectives
• Conclusions

4
The System Overview
5
Histogram Segmentation

• Automatically Picking a Threshold
• Otsu thresholding method
• - minimization of the weighted within-class
  variance / maximization of the inter-class
  variance
• Pal thresholding method
• - concept of cross-entropy maximization

6
Histogram SegmentationResults

• works well on simple images

Original image
Otsu
Pal
7
K-means

• k-means clustering method of cluster analysis
  -gt partitions n observations into k clusters in
  which each observation belongs to the cluster
  with the nearest mean
• set of observations (x1, x2, , xn) -gt partition
  the n observations into k sets (k lt n)
• Basic steps
• -gt -gt -gt

8
K-means Results

• automatic computing of number of
  classes/clusters -gt peaks histogram detection

Original image
Output image
9
Expectation Maximization

• EM algorithm maintains probabilistic assignments
  to clusters, instead of deterministic
  assignments
• E step assign points to the model that fits it
  best
• M step update the parameters of the models using
  only points assigned to it

10
Expectation Maximization Results

• automatic computing of number of
  classes/clusters -gt peaks histogram detection

11
Illuminant invariant images

• RGB -gt 2D log-chromaticity co-ordinates
• r log(R) log(G)
• b log(B) log(G)
• the r and b co-ordinates varies when illumination
  changes
• the pair (r,b) for a single surface viewed under
  many different lights - a line in the
  chromaticity space
• projecting orthogonally to this line results in a
  1D value which is invariant to illumination
• by subtracting from the grayscale image the
  illuminant invariant, we obtain a perfect mask of
  the shadow

12
Shadow Removal

• Illumination recovery
• recover the illuminated intensity at a shadowed
  pixel -estimate the four parameters of the affine
  model
• two strips of pixels one inside the shadowed
  region, and the other outside the region
• S -gt shadowed set of pixels
• L -gt illuminated set of pixels
• and denote the mean colors of pixels
  from S and L
• and denote the standard deviations

13
Shadow Removal

• Inpainting
• the patch lies on the continuation of an image
  edge, the most likely best matches will lie along
  the same (or a similarly colored) edge
• the algorithm is divided in 3 steps
• compute patch priorities
• propagate texture and structure information
• update confidence values.

14
Illuminant invariant images Shadow removal
Results
15
Future Perspectives
16
Future Perspectives
17
Future Perspectives

• To be in contact with all participants of SSIP

18
Conclusions

• The proposed method is fully automatic (no user
  interaction)
• Several methods of shadow detecting have been
  applied and good reasults have been reached
• The methods of shadow removal should be improved
  for complex images

19
Thank you for your attention !