Inter-frame prediction with fast weighted low-rank matrix approximation

Zhi Long Huang, Hsu-Feng Hsiao

Research output: Contribution to journalArticlepeer-review

1 Scopus citations


In the field of video coding, inter-frame prediction plays an important role in improving compression efficiency. The improved efficiency is achieved by finding predictors for video blocks such that the residual data can be close to zero as much as possible. For recent video coding standards, motion vectors are required for a decoder to locate the predictors during video reconstruction. Block matching algorithms are usually utilized in the stage of motion estimation to find such motion vectors. For decoder-side motion derivation, proper templates are defined and template matching algorithms are used to produce a predictor for each block such that the overhead of embedding coded motion vectors in bit-stream can be avoided. However, the conventional criteria of either block matching or template matching algorithms may lead to the generation of worse predictors. To enhance coding efficiency, a fast weighted low-rank matrix approximation approach to deriving decoder-side motion vectors for inter frame video coding is proposed in this paper. The proposed method first finds the dominating block candidates and their corresponding importance factors. Then, finding a predictor for each block is treated as a weighted low-rank matrix approximation problem, which is solved by the proposed column-repetition approach. Together with mode decision, the coder can switch to a better mode between the motion compensation by using either block matching or the proposed template matching scheme.

Original languageEnglish
Pages (from-to)9-16
Number of pages8
JournalInternational Journal of Electronics and Telecommunications
Issue number1
StatePublished - Mar 2013


  • Block matching
  • Inter-frame prediction
  • Low-rank matrix approximation
  • Template matching
  • Weighted low-rank matrix approximation


Dive into the research topics of 'Inter-frame prediction with fast weighted low-rank matrix approximation'. Together they form a unique fingerprint.

Cite this