Real-Time Low-Power Super-Resolution on the Edge: A Comparative Study of Pruned GANs and Post-Training Quantization

@article{1722518,
      author = {Hiroaki Tan, Sofia Marchetti, Rahul Bansal, Kwame Osei},
      title = {Real-Time Low-Power Super-Resolution on the Edge: A Comparative Study of Pruned GANs and Post-Training Quantization},
      journal = {Iconic Research And Engineering Journals},
      year = {2022},
      volume = {6},
      number = {3},
      pages = {397-402},
      issn = {2456-8880},
      url = {https://www.irejournals.com/formatedpaper/1722518.pdf},
      abstract = {Single-image super-resolution is increasingly demanded on mobile and embedded platforms, yet generative super-resolution networks are notoriously heavy. This paper studies how structured channel pruning combined with post-training quantization (PTQ) can deliver real-time 4x super-resolution within the power envelope of edge accelerators. We prune three GAN generators to 40% of their original channel width and quantize them to INT8 and INT4, then benchmark reconstruction fidelity, latency, and power on three representative devices. Our results show that pruning-then-PTQ preserves most perceptual quality while cutting latency by up to 47% and energy by roughly a third.},
      month = {September},
  }