Image and Video Error Concealment

1
Image and Video Error Concealment

• By Narek Zohrabyan

2
Imperfect medium

• Video transmission susceptible to data loss.
• Noise.
• Burst errors.
• Transmitted compressed images are not inherently
  error resilient.
• Issue more restrictive in these areas
• Real-time video communication.
• Video broadcasting.
• Internet video delivery.
• By exploiting statistical information, and
  manipulating spatial and frequency dependencies,
  error can be concealed.

3
Justification

• Critical information delivery is important.
• Effected industries are
• Medical
• Commercial
• Aerospace
• Defense
• Science research and exploration.

4
Background information

• Typical image in raw format is very big.
• Video (sequence of images) is astronomically
  bigger.

DSL 1.5Mbps, 1 Hour film 75GB where one DVD
disc is about 4.7GB
5
Data Compacting

• Images have redundant data in the spatial domain.
• Neighboring pixels have little variations.
• Video has redundant data in the temporal domain.
• Adjacent frames contain similar values.
• Compressing images and video.
• DCT (Discrete Cosine Transform)
• Wavelet Transform

Bitstream 011101
T
Q
Encoding
Transform coefficients
Input image
Quantization
Transform
6
DCT

• Pixel values are transformed into coefficients in
  the frequency domain.
• Separating low frequency (important) data from
  high frequency (unimportant) data.
• Pixel values are de-correlated.

7
Wavelet Transform

• Image is decomposed in sub-bands.
• LL, LH, HL, HH.
• Advantage over DCT
• Coefficients map spatial information.
• Functions better in low data rate applications
• Better performance with error concealment
• Progressive coding
• Bitstream can be truncated arbitrarily for
  reconstruction
• Up to 3db improvement

8
Wavelet Transform
1-layer transform
2-layer transform
Original Image
Wavelet Coefficients
9
Embedded Zero-tree Wavelet coding (EZW)

• Each parent has four children with the exception
  of the lowest band.
• Descendents and ancestors are included in the
  tree.
• Coefficients are flagged as either significant,
  isolated zero, negative/positive significant and
  zero-tree root.

From Usevitch (IEEE Sig.Proc. Mag. 9/01)
10
DCT vs. EZW
From Christopoulos (IEEE Trans. on CE 11/00)
11
Block Packetization

• Significance maps are packetized sequentially.

P1
P2
P3
Pn
12
Dispersed Packetization

• Interleaving packets allows for better
  performance in case one packet is lost.

2
3
1
6
4
5
7
8
13
Passive Error Concealment

• Using local coefficients, the decoder can
  interpolated missing pixels.
• Bilinear interpolation
• Locally adaptive passive error concealment

14
Results

• (a) original compressed image (PSNR 32.18 dB)
• (b) loss of packet 8
• (c) Bilinear interpolation (PSNR 28.35 dB)
• (e) Locally adaptive passive error concealment
  (PSNR 28.86 dB)

15
LabVIEW examples