Title: Perceptual Image Filtering with Connected Operators and Image Inpainting Department of Signal Theory and Communications Technical University of Catalonia (UPC) Barcelona, Spain
1Perceptual Image Filtering with Connected
Operators and Image Inpainting Department of
Signal Theory and CommunicationsTechnical
University of Catalonia (UPC)Barcelona, Spain
- Mariella Dimiccoli and Philippe Salembier
2Objectives
Introduction ??
- Objective
- Improvement of filtering based on Connected
Operators
Connected Operators
3Objectives
Introduction ??
- Objective
- Improvement of filtering based on Connected
Operators
Image Inpainting
Connected Operators
through Image Inpainting
4Objectives
Introduction ??
- Presentation Outline
- Introduction
- Connected Operators
- Image Inpainting
- Proposed Approach
- Experimental Results
- Conclusions
5Objectives
Connected Operators ????
- Presentation Outline
- Introduction
- Connected Operators
- Image Inpainting
- Proposed Approach
- Experimental Results
- Conclusions
6Objectives
Connected Operators ????
Introduction
- Definition Flat zones that do not fulfill a
given simplification criterion are removed and
merged into a single one - Property Cannot create new contours nor shift
existing ones
- Construction strategies
- Reconstruction
- Region-tree pruning
7Objectives
Connected Operators ????
Connected Operators by Reconstruction
- Relies on a pixel-based representation of the
image - is the reference image
- is the marker image
- is a generic reconstruction (leveling)
A connected operator by reconstruction can be
written as
where is a transformation that defines the
simplification effect
8Objectives
Connected Operators ????
Connected Operators by Region Tree Pruning
Relies on a region-based representation of the
image
- The general approach involves three steps
- Tree-construction
- Tree root represents the image support
- Tree nodes represent image regions
- Tree branches represent inclusion relationships
between nodes - Tree-pruning
- Image construction
9Objectives
Connected Operators ????
Perceptual Analysis
- Drawback Removed regions may be still
perceptible in the filtered image as single flat
zones
Size simplification Opening by reconstruction
of erosion
Proposed Solution Use inpainting to estimate
pixel values of flat zones that have been removed
and merged
10Objectives
Image Inpainting ??
- Presentation Outline
- Introduction
- Connected Operators
- Image Inpainting
- Proposed Approach
- Experimental Results
- Conclusions
11Objectives
Image Inpainting ??
Introduction
What is Image Inpainting? Restoring tool allowing
the recovery of a given part of an image based on
the surrounding information
- Classification
- Textural (textural exemplares)
- Structural (contour information)
12Objectives
Image Inpainting ??
A state-of-Art Inpainting Algorithm
Variational Method Join interpolation of
isothopes and gray levels (Ballester et all.,
2001)
Basic Idea For each possible continuation, the
angular change of the normal along the whole
contour in is minimized
?
B
Masked
Inpainted
Original
Study of the convergence For each level set the
stopping criterion is defined the by a
thresholding on the slope of the energy function
13Objectives
Proposed Approach ??
- Presentation Outline
- Introduction
- Connected Operators
- Image Inpainting
- Proposed Approach
- Experimental Results
- Conclusions
14Objectives
Proposed Approach ??
- Basic idea Use inpainting to estimate pixel
values of flat zones - that have been removed and merged
15Objectives
Proposed Approach ??
- Basic idea Use inpainting to estimate pixel
values of flat zones - that have been removed and merged
Which flat zone need to be estimated by
inpainting? Only the perceptually most important
removed regions
16Objectives
Proposed Approach ??
17Objectives
Proposed Approach ??
18Objectives
Proposed Approach ??
19Objectives
Proposed Approach ??
20Objectives
Proposed Approach ??
21Objectives
Proposed Approach ??
22Objectives
Proposed Approach ??
23Objectives
Proposed Approach ??
24Objectives
Proposed Approach ??
25Objectives
Proposed Approach ??
26Objectives
Proposed Approach ??
27Objectives
Experimental Results ?????
- Presentation Outline
- Introduction
- Connected Operators
- Image Inpainting
- Proposed Approach
- Experimental Results
- Conclusions
28Objectives
Experimental Results ?????
Size criterion
Size criterion is defined as allows
It allows the removal of those flat zones
that represent a maxima of size inferior to the
size of the structuring element
29Objectives
Experimental Results ?????
Contrast criterion
Contrast criterion is defined as follows
It allows the removal of those flat zones that
have a contrast inferior to
Original
Filtered
Mask
Result
30Objectives
Experimental Results ?????
Motion criterion
Motion criterion allows the removal of those flat
zones that do not undergo a given motion
Original
Filtered
Mask
Result
31Objectives
Experimental Results ?????
Complexity and contrast criteria
Original
Reference
32Objectives
Experimental Results ?????
Complexity and contrast criteria
Complexity Criterion allows the removal of those
flat zones that have a big ratio between
perimeter and area
Original
33Objectives
Experimental Results ?????
Complexity and contrast criteria
Complexity Criterion allows the removal of those
flat zones that have a big ratio between
perimeter and area
Original
Contrast Oriented Connected Operator
Complexity Oriented Connected Operator
IA
IB
ID
IE
-
IC
34Objectives
Experimental Results ?????
Complexity and contrast criteria
Original
35Objectives
Experimental Results ?????
Complexity and contrast criteria
Original
36Objectives
Experimental Results ?????
Complexity and contrast criteria
Original
37Objectives
Experimental Results ?????
Complexity and contrast criteria
Result
Original
38Objectives
Experimental Results ?????
Complexity and contrast criteria
Original image detail
Original
Filtered
Result image detail
Mask
Result
39Objectives
Experimental Results ?????
- Given a reference and a distorted image
- The PSNR (Peak Signal to Noise Ratio) quantifies
objectively the distortion but it is not related
with the perceptual visual appearance. - The SSIM (Structural Similarity Image Measure)
quantifies the perceptual similarity
independently on local contrast and local
luminance changes - The SSIM is equal to 1 for two identical images
Average
Median
Type of filter PSNR SSIM
Average (5x5) 19.8 0.06
Median (5x5) 23.2 0.87
Connected Operator 21.9 0.91
Proposed Approach 39.2 0.99
Connected Operator
Proposed Approach
40Objectives
Conclusions ?
- Presentation Outline
- Introduction
- Connected Operators
- Image Inpainting
- Proposed Approach
- Experimental Results
- Conclusions
41Objectives
Conclusions ?
- Summarizing..
- A filtering strategy combining Connected
operators and Image Inpainting has been
presented. - Based on experimental results, the proposed
approach is - Efficient Outperforms objectively other
filtering strategies - Flexible Any connected operator can be used
involving also semantic criteria - Future work will be devoted to improve the
current method to select the regions to be
inpainted
42Perceptual Image Filtering with Connected
Operators and Image Inpainting
Mariella Dimiccoli and Philippe Salembier
43Perceptual Image Filtering with Connected
Operators and Image Inpainting
Mariella Dimiccoli and Philippe Salembier