Transcript Artifact and Textured region Detection

Artifact and Textured
region Detection
- Vishal Bangard
Outline
 Need for artifact and textured region
detection
 Aim of the project
 Techniques used in the imaging world
 Approaches used
 Results
 Conclusion
Why do artifact detection ?
 A lot of transformations lead to artifacts
 Few of them lead to loss in texture
 Main goal – repair/ replace the loss in texture
using texture from adjacent regions
 There are existing methods for replicating
texture
 Not many existing methods for detecting
regions where there is texture loss
Left image compressed at 94%
Image: Barbara
The encircled areas show loss in
texture due to compression
Aim of the project
 Very subjective – depends a lot on prior
experience and knowledge
 Complete automated detection is very hard
because of the subjective nature of the
problem
 Aim of this project: To locate regions near
textured regions which may have been
subject to texture loss
Techniques that work well
 The topics covered under this project are very
subjective – hence the title of this slide is
‘techniques that work well’ as against ‘state of
the art approaches’
 Detection of textured regions


Gabor filters
Difference of offset Gaussians
 Segmentation


Lots and lots of them
(e.g. thresholding, clustering methods such as
k-means and fuzzy c-means, connected
components, region growing, etc.)
Again, very subjective
Approaches taken
 Analyze wavelet block
decomposition for a
change in the high
frequency region


Low resolution
Higher miss probability
The image Barbara compressed at 94%
Solid black areas give regions with
texture. Black lines are parts of
edges
Edge map of left image overlaid on
the compressed image
 Use of Gabor coefficients as they are fairly
reliable in detecting textured regions (found at
six different orientations and four different
scales)

Transformed image is adaptively thresholded
to minimize the inter-class variance (Ref: N. Otsu, "A
Threshold Selection Method from Gray-Level Histograms," IEEE Transactions on Systems,
Man, and Cybernetics, vol. 9, no. 1, pp. 62-66, 1979.)

Magnitude and phase information is used
collectively to locate regions of high texture
Pictures speak a thousand words
White regions give borders and
textured areas
An edge map of the left image
overlaid on the compressed image
Image: Baboon
Left image compressed at 90% using
Elecard (version of JPEG 2000
coding software)
White regions give regions of high
textures
An edge map of the left image
overlaid on the compressed image
Conclusion
 The system performs well in detecting
textured regions (Statistics still need to be
calculated)
 Needs to be extended to color and non-
square images
 Can also be used for texture identification
 Can be combined with existing methods to
repair texture loss