.png){kind=icon}
.png){kind=icon}
Return
Deep learning image compression with multi-channel tANS coding and hardware deployment
.png){kind=icon}
PRE
AI
DOI：10.1007/s11554-025-01795-8.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Deep learning-based image compression outperforms traditional methods in coding efficiency, but its computational complexity hinders real-time deployment on embedded devices. This paper proposes a heterogeneous computing system combining GPU-accelerated inference and CPU-accelerated entropy coding via lookup tables, breaking performance bottlenecks through algorithm-hardware co-design. After GPU acceleration, entropy coding becomes the dominant bottleneck (73% of runtime). To address this, we introduce three key innovations: replacing rANS with tANS encoding, converting dynamic computations into static table lookups, reducing encoding latency; a cache-friendly tANS coding scheme for the 192-channel network outputs, minimizing access latency; an out-of-range symbol encoding method, ensuring lossless and efficient compression. Experiments demonstrate that under high compression ratios, compared with traditional rANS, tANS reduces latency by 77%, with a compression ratio loss of 12.6% while still ensuring image compression quality higher than JPEG2000.
Keywords:
Deep learning image compression
tANS coding
Entropy coding
Cache-friendly
GPU
Journal
IF:
3
Papers:
371
Citations:
2.0K