Outline Introduction Inter Mode Decision in Transcoding Using SVM - - PDF document

1

1

Efficient Inter Mode Decision for H.263 to H.264 Video Transcoding Using Support Vector Machines Speaker: Prof. Wan-Chi Siu

Xuan Jing*, Wan-Chi Siu*, Lap-Pui Chau+ and A.G. Constantinides**

*Department of Electronic and Information Engineering, Hong Kong Polytechnic University

+School of Electrical and Electronic Engineering, Nanyang Technological University, Singapore

**Department of Electrical and Electronic Engineering, Imperial College London

IEEE International Symposium on Circuits and Systems (ISCAS’2009) 24-27 May 2009, Taipei, Taiwan

2

Outline

 Introduction  Inter Mode Decision in Transcoding Using SVM  Experimental Results  Conclusions

2

3

Introduction



H.264 Interframe Mode Decision



Inter block partition



Candidate MB modes in P-frame

SKIP, Intra-4 4, Intra-16 16, mode Inter-16 16, Inter-16 8, Inter-8 16, Inter-8 8            

4

Introduction (cont.)



Lagrangian RD cost function in H.264:



Objectives of proposed mode decision method for transcoding



Making use of the information from H.263 decoding process to predict the H.264 Inter block coding mode for each MB.



SVM-based mode prediction to early reject unlikely modes in

• rder to achieve complexity saving

The basic idea of fast mode decision in H.264 is to select the coding mode that achieves the best RD performance without searching all the modes, leading to complexity reduction.

mode mode

( , ,mode| , ) ( , ,mode| ) ( , ,mode| ) J s c QP SSD s c QP R s c QP     

3

5

Inter Mode Decision in Transcoding Using SVM



Proposed H.263 to H.264 transcoder with fast Inter mode decision

6

Inter Mode Decision in Transcoding Using SVM (cont.)



Support vector machines (SVM)



Based on a training set of examples with class labels and several attributes, the SVM constructs a decision rule which can be used predict the class lables of new input data given only attributes.



In H.263 to H.264 transcoding:

• - Class labels correspond to the optimal MB modes in H.264
• - Attributes correspond to extracted MB features from H.263

4

7



Four extracted features from H.263 as attributes:

Inter Mode Decision in Transcoding Using SVM (cont.)

MB mode in H.263 {Skip, Intra, Inter16x16, Inter8x8} Mean value of the MB residue Variance of the MB residue The number of non-zero AC coefficients of the MB, i.e. nzcoef



Recall the SVM problem formulation: (e.g. binary classification)



Given a training set of and a vector such that

(to minimize the complexity (large margin) and training error)

( , )

i i

y x

y

{1, 1}

i

y  

, , 1

1 min 2 s.t. ( ( ) ) 1 , 0, 1,....

N T i b i T i i i i

C y b i N    



     



w ξ

w w w x

C: Error penalty parameter N: total number of training data w is the normal vector perpendicular to the hyperplane and i are slack variables

Some terms 8

 SVM offline training process

Inter Mode Decision in Transcoding Using SVM (cont.)

* C.-C. Chang and C.-J. Lin, LIBSVM: a library for support vector machines. Software available at http://www.csie.ntu.edu.tw/~lin/libsvm



Based on free libSVM* software



Four typical test sequences (Foreman, Stefan, Salesman, News) were firstly encoded and decoded by H.263 codec and then re-encoded by H.264 encoder



Use RBF (radial basis function) kernel in SVM



Use cross-validation to find the best penalty parameter, C, and the kernel parameter

Some terms

Cross-Validation

5

9

Inter Mode Decision in Transcoding Using SVM (cont.)

 SVM training results



A few types of classification considered:

• - Two-class classification {Skip, non-Skip}
• - Three-class classification {Skip, Intra, Inter}
• - Four-class classification {Skip, Intra, InterL, P8x8}

(InterL includes Inter16x16, Inter16x8 and Inter8x16)



Prediction accuracy of SVM classifier

More than 80% accuracy is achieved. In our proposed transcoder, a few possible modes will be further added to the prediction mode list of the four-class SVM classifier output. 10



Summary of the proposed Inter mode decision for transcoding

Inter Mode Decision in Transcoding Using SVM (cont.)

6

11

Experimental Results

• Comparison of Inter mode decision algorithms for transcoding

The H.264 with full search mode decision is used as benchmark for comparison

Average -0.02 0.37 -53.14

• 0.03 0.61 -73.50

Alg in Literature 12

Experimental Results (cont.)

• R-D performance comparison

Mother&Daughter Coastguard

7

13

Conclusions

 Proposed fast Inter mode decision for H.263 to

H.264 video transcoding.

 Using selected features from H.263 to predict

possible candidate modes in H.264 inter mode decision.

 SVM-based classifier is used for mode

prediction

 Up to 77% complexity reduction is achieved

with similar RD performance.

• Adv. & Disadv.

14

Thank You!