Energy-Efficient On-Device Image Restoration: Benchmarking Compact GANs and Mixed-Precision Quantization for Embedded Vision

@article{1722519,
      author = {Devang Kulkarni, Lena Havlik, Priyanka Menon},
      title = {Energy-Efficient On-Device Image Restoration: Benchmarking Compact GANs and Mixed-Precision Quantization for Embedded Vision},
      journal = {Iconic Research And Engineering Journals},
      year = {2022},
      volume = {6},
      number = {4},
      pages = {233-238},
      issn = {2456-8880},
      url = {https://www.irejournals.com/formatedpaper/1722519.pdf},
      abstract = {On-device image restoration must remove noise and blur while respecting tight energy budgets. We benchmark three compact restoration GANs under mixed-precision quantization, where weights are held at INT8 and the most sensitive activations remain in FP16. Across denoising and deblurring benchmarks, the mixed-precision scheme recovers nearly full-precision fidelity while lowering per-frame energy to as little as 96 mJ, establishing a favourable accuracy-per-joule frontier for embedded vision.},
      month = {October},
  }