Fast and energy-efficient approximate motion estimation architecture for real-time 4 K UHD processing
| Autor: | Murilo Perleberg, Vladimir Afonso, Luciano Agostini, Nuno Roma, Roger Porto, Bruno Zatt, Marcelo Porto |
|---|---|
| Rok vydání: | 2020 |
| Předmět: |
Pixel
Computer science Process (computing) 020207 software engineering 02 engineering and technology Frame rate Computer engineering Motion estimation 0202 electrical engineering electronic engineering information engineering 020201 artificial intelligence & image processing Encoder Throughput (business) Information Systems Integer (computer science) Efficient energy use |
| Zdroj: | Journal of Real-Time Image Processing. 18:723-737 |
| ISSN: | 1861-8219 1861-8200 |
| DOI: | 10.1007/s11554-020-01014-6 |
| Popis: | Approximate computing techniques exploit the characteristics of error-tolerant applications either to provide faster implementations of their computational structures or to achieve substantial improvements in terms of energy efficiency. In video encoding, the motion estimation (ME) stage, including the Integer ME (IME) and the Fractional ME (FME) steps, is the most computational intensive task and it is highly resilient to controlled losses of accuracy. In accordance, this article proposes the exploitation of approximate computing techniques to implement energy efficient dedicated hardware structures targeting the motion estimation stage of current video encoders. The designed ME architecture supports IME and FME and is able to real-time process 4 K UHD videos (3840 × 2160 pixels) at 30 frames per second, while dissipating 108.92 mW. When running at its maximum operation frequency, the architecture can process 8 K UHD videos (7680 × 4320 pixels) at 120 frames per second. The solution described in this article presents the highest throughput and the highest energy efficiency among all state-of-the-art compared works, showing that the use of approximate computing is a promising solution when implementing video encoders in dedicated hardware. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|
Full text from SpringerLink