A DCT-Domain Video Transcoder for Spatial Resolution Downconversion

Document Sample
A DCT-Domain Video Transcoder for Spatial Resolution Downconversion Powered By Docstoc
					A DCT-Domain Video Transcoder for
 Spatial Resolution Downconversion
 Yuh-Reuy Lee, Chia-Wen Lin and Cheng-Chien Kao
      CS@National Chung Cheng University

                            Presented by Bin Yu
                 Outline
• Motivation
• Transcoder Architecture
• Spatial Downscaling
  – Sub-sampling
  – DCT Decimation
• Motion Vector Composition
• Conclusion
                Motivation
• Universal Multimedia Access
  – Heterogeneous devices/connections in the
    network



                  Spatial Downscaling
           Spatial Resolution Downscaling Illustration
  3   4    3   4    6   8   3    6
  7   2    7   2    5   9   6    8           2   1   3     -1       8     9   6    5
  9   10   9   10   5   6   3    8           4   2   5     1        9    11   11   9
  4   8    4   8    6   4   2    3       +   7   8   2     6    =   16   18   5    14
  5   7    5   7    9   1   3    7           6   4   2     4        10   12   4    7
  2   3    2   3    5   7   5    4             Motion                Resulting
  3   4    3   4    6   8   3    6           Vectors and            Macroblocks
                                             Differences
  7   2    7   2    5   9   6    8
      Reference Macroblocks

Downscaling
           4   4    7   6
                                Point to ?
           8   8    5   4                                       ?
                                                 2 2                     9 8
           4   4    5   5            +           6 3
                                                                =        14 8
           4   4    7   6
Cascaded DCT-domain Transcoder Architecture
Cascaded DCT-domain Spatial Resolution
 Downscaling Transcoder Architecture
           Spatial Downscaling scheme 1 – Sub-sampling
     Pixel Domain Sub-sampling                            DCT Domain Equivalent

     b1         b2                                    4
     8x8        8x8                  b1 b2
                                     4x4 4x4
                                               b   hi * bi * g i
                                                     i 1
                                     b3 b4                       4
     b3         b4                   4x4 4x4   DCT (b)   DCT (hi ) * DCT (bi ) * DCT ( g i )
     8x8        8x8                                             i 1

                                          b                      4
                                                                H i * Bi * Gi
                                                                i 1
Example:
      .5   .5   0     0    0    0    0    0
       0   0    .5    .5   0    0    0    0
       0   0    0     0    .5   .5   0    0
       0   0    0     0    0    0    .5   .5                                                b1
h1                                             h1                 b1              g1
                                                                                        =
                                                                                            4x4
       0   0    0     0    0    0    0    0               *               *
                                               8x8                8x8             8x8
       0   0    0     0    0    0    0    0
       0   0    0     0    0    0    0    0
       0   0    0     0    0    0    0    0

g1 = h1’
          Spatial Downscaling scheme 2 – DCT
                      Decimation
 • Key observation: hi are sparse, but DCT(hi) are not!


   B1         B2
                          ˆ
                          B1         ˆ
                                     B2                                       8x8 DCT
   8x8        8x8                                                     ˆ ˆ
                                                                      b1 b2
                                                 4x4 IDCT
                          ˆ
                          B3         ˆ
                                     B4                               ˆ ˆ
   B3         B4                                                      b3 b4
   8x8        8x8
                                                                        ˆ
                                                                        b                    ˆ
                                                                                             B
                           ˆ ˆ
                         b1 b2  TL                    ˆ        ˆ
                                                     T4t B1T4 T4t B2T4  TL 
ˆ     ˆ
B  TbT  [TL TR ]
           t
                                    [TL TR ] t                       
                           ˆ b  TR 
                              ˆ
                         b3 4                           ˆ T T t B T  TR 
                                                                    ˆ
                                                     T4 B3 4 4 4 4   
                                                    
            ˆ                       ˆ                        ˆ                       ˆ
 (TLT4t ) B1 (TLT4t ) t  (TLT4t ) B1 (TRT4t ) t  (TRT4t ) B1 (TLT4t )t  (TRT4t ) B1 (TRT4t )t
   Motion Vector Composition
• Composing one new MV from 4 old MVs
  – Majority voting
  – Averaging
  – Median filtering: the one MV has the least
    distance to all other MVs             4


• Activity-weighted Median
                                       di      || v  v
                                              j 1, j  i
                                                            i   j   ||

  – Use squared sum of all coefficients as the
                                      4
    activity of a MB          1
                       di          ||i vi  v j ||
                             ACTi j 1, j
                  Conclusion
• Contribution
  – A new dimension of stream adaptation
  – A suite of DCT-compressed domain algorithms
• Future work
  – Real-world implementation
  – Going for HDTV
     • HDTV 1920x1088
     • PC/Laptop 960x544
     • iPAQ 240x136