Pixel Interpolation

1
Pixel Interpolation

• By Mieng Phu
• Supervisor Peter Tischer

2
Outline

• Pixel Interpolation and Background
• Scenarios
• Interpolation Techniques
• Test Data
• Results and Discussions
• Future Work
• Summary

3
Pixel Interpolation

• Predicting/interpolating missing values occurs in
  many areas of image processing, especially in
  lossless image coding.
• The idea of this project is to take the
  prediction techniques from lossless image coding
  and apply them to other area of image/video
  processing.

4
Scenarios in Video/Image Processing

• Deinterlacing within a single field (spatial)
• Deinterlacing in two fields (spatial and
  temporal)
• Deinterlacing in three fields (spatial and
  temporal)
• Convert from SDTV to HDTV (Magnification)

5
X- known value
? - unknown value
Scenario One (1 field)
Scenario Two (2 fields)
6
Scenario Three (3 fields)
7
Terminology

• Neighboring pixels

8
Interpolation Techniques(1)

• Scenario One Predictors (Prediction within a
  single field or on still image).
• Line doubling
• Averaging, e.g. (NS)/2, (NWNNESWSE)/2
• Median filter.
• Pseudo Median (PMED) - H-Shape, A-Shape,
  Adaptive.
• ELA, Adaptive ELA, ELA (Tao Chen).
• Some of my proposed algorithms.

9
Interpolation Techniques(2)

• Edge Line Averaging (ELA)
• Adaptive - ELA (A-ELA)
• The concept is the same as ELA, but it uses a
  unique way to detect a horizontal edges.

10
Interpolation Techniques(3)

• ELA Chen, Henry, et al.
• Used two additional measurements to determine the
  direction correlations.
• Hence, good predictor for a 630 edges from the
  horizontal.

11
Interpolation Techniques(4)

• Median Filter
• Median10,10,10,10,100,100
• Pseudo Median (by definition)

Segment 1
Segment 2
PMED a, b, c, d, e, f 0.5 x max min of each
sub window 0.5 x min max of each sub window
12
Interpolation Techniques(5)

• For scenario two and three predictors
• Interpolation techniques in situation scenario
  one can be generalized in situation two and three.

? h ? (b h e)/3 ? median (b, h, e,)
? PMED a, b, c, d, e, f, h
13
? (h k)/2 ? Median h, k, (be)/2 ? PMED
a, b, c, d, e, f, h, k
14
Test Data

• Standard natural Images
• Synthetic Images
• Lines with different orientations
• Different textures
• Video sequences
• Different speeds of motion between fields.
• Textures edges, lines etc.

15
Scenario1 Results(1)
16
Scenario1 Results(2)

• H-Shaped PMED
• Best predictor overall
• Superb in detection of
• vertical edges.

H-Shaped PMED a, b, c, d, e, f 0.5 x
maxmina,b,c,,mind,e,f,minb,e 0.5 x
minmaxa,b,c,maxd,e,f,maxb,e
17
Scenario1 Results(3)

• LD - 21.61 dB
• ELA - 19.26 dB
• A-ELA - 19.16 dB
• ELA (CHEN) - 24.99 dB
• But even better results
• H-Shaped - 25.27 dB
• CHEN PMED - 25.22 dB

18
Scenario1 Results(4)

• CHEN PMED is a combination of the PMED and ELA
  (Chen).
• It use ELA(Chen) to select the PMED subwindows.
• Like ELA(Chen), it predict well for edges at an
  orientation of 630.
• This algorithm perform better than ELA(Chen) in
  wide range of images.

19
Scenario1 Results(5)

• The median a, b, c, d, e, f, (be)/2
• Can have the
• median a, b, c, d, e, f, (be)/2, (ab)/2,
  (bc)/2, (de)/2, (ef)/2
• Furthermore, you can further subdivide into
  half-pel, quarter-pel, or even further.
• This approach is proven to be better than the
  generic median filter. But more computation.

20
Scenario1 Results(6)

• A-ELA can detect horizontal lines incredible
  well.
• A-ELA 39.7 dB
• H-Shaped 25.7 dB
• ELA(CHEN) - 25.7dB
• A-ELA - 13.15 dB
• H-Shaped - 8.47 dB
• ELA(CHEN) 8.45 dB

21
Scenario1 Results(7)
Football (64 frames). 352x240
Flower Garden (64 frames). 352x288
Akiyo (10 frames). 352x288
22
(No Transcript)
23
Conclusions

• ELA (Chen) is good for detection of diagonal
  edges.
• ELA (Chen) can be improved by using PMED or
  Median filter.
• H-Shaped PMED can detect the vertical edges well,
  and perform best overall.
• A-ELA is really good at detect the horizontal
  lines.
• Median filter can be improved by using the right
  group of pixels.
• Average 2, give good results and for little
  computation.

24
Future Work

• Potential algorithms in scenario one can be
  improved.
• Combine H-Shaped, ELA(CHEN) and A-ELA together,
  to form the best predictor.
• More generalization can be made on two and three
  fields.
• Magnifications

25
Questions?