A Novel Video Compression Approach Based on Underdetermined Blind Source Separation

Jing Liu, Fei Qiao, Qi Wei, Huazhong YangarXiv 2012 · arXiv

Abstract

As digital video demand continues to grow, traditional compression standards are approaching a performance plateau. This paper presents UBSSVC (Underdetermined Blind Source Separation based Video Compression), a novel framework developed by researchers at Tsinghua University. By intentionally mixing multiple video frames into a smaller subset before standard encoding and subsequently "unmixing" them at the decoder using Sparse Component Analysis (SCA), the proposed method achieves compression ratios up to 12 times higher than H.264 for specific video types. The approach is particularly well-suited for low-power encoding environments such as remote surveillance.

Introduction

The proliferation of digital video in surveillance and communication has created a persistent gap between data volume and available bandwidth. While established codecs like H.264 and MPEG-2 are highly efficient, they eventually encounter a "compression wall." To overcome this, the authors propose treating video frames as signals that can be mathematically combined and separated. By utilizing underdetermined systems—where the number of output frames is less than the number of input frames—the framework pushes the boundaries of data reduction beyond conventional limits.

The UBSSVC Framework

The UBSSVC system operates through a three-stage pipeline designed to shift computational complexity from the encoder to the decoder.

Block diagram of the UBSSVC framework showing the mixing process, standard compression, and separating process.

1. Factitious Mixing (Encoder)

At the encoder, $n$ raw video frames are multiplied by a predetermined $m \times n$ mixing matrix ($A$), where $m < n$. This "factitious mixing" reduces the frame count before any standard compression occurs. Because the mixing matrix is known at the decoder, the recovery process is significantly more stable than traditional blind source separation.

2. Standard Compression

The resulting $m$ mixed frames are processed using off-the-shelf codecs such as H.264 or MPEG-2. This ensures compatibility with existing infrastructure while benefiting from the initial data reduction.

3. Sparse Recovery (Decoder)

The decoder performs the inverse operation. To solve the underdetermined system (more unknowns than equations), the framework leverages the sparsity of video signals in the transform domain. Frames are converted using a 2-D discrete Haar wavelet transform, where most coefficients are near zero. Sparse Component Analysis (SCA) then identifies the most likely original signals by navigating the hyperplanes defined by the mixing matrix.

Experimental Evaluation

The researchers tested UBSSVC using four standard video sequences representing various motion levels: Hall (surveillance, low motion), Container (slow motion), Foreman (moderate motion), and Football (high motion).

The four standard video test sequences used in the study: Hall (top-left), Container (top-right), Foreman (bottom-left), and Football (bottom-right).

Performance Gains

The integration of UBSSVC provided substantial improvements over standalone codecs:

  • MPEG-2: Compression ratios improved by approximately 33% across all sequences.
  • H.264: The gains were most pronounced in low-motion scenarios. For the "Hall" sequence, the file size was reduced from 11.6 KB to 3.89 KB, representing a 4x to 12x improvement in compression ratio depending on the specific configuration.

Discussion and Limitations

While UBSSVC offers impressive data reduction, its performance is sensitive to temporal dynamics. High-motion sequences like "Football" resulted in lower Peak Signal-to-Noise Ratios (PSNR), as rapid changes challenge the sparsity assumptions required for accurate SCA recovery. Consequently, the current iteration is best suited for surveillance and monitoring applications where static backgrounds are common and extreme compression is prioritized over high-fidelity cinematic quality.

Conclusion

UBSSVC introduces a promising paradigm for video compression by combining signal processing theory with standard codecs. Its low-complexity encoding makes it an ideal candidate for battery-powered edge devices. Future work will focus on optimizing mixing matrices and refining recovery algorithms to better handle high-motion content and improve overall visual fidelity.

Summary updated

BibTeX
@misc{liu2012novel,
  author = {Jing Liu and Fei Qiao and Qi Wei and Huazhong Yang},
  title = {{A Novel Video Compression Approach Based on Underdetermined Blind Source Separation}},
  year = {2012},
  eprint = {1205.4572},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1205.4572},
}
Figure