# Real Image Denoising with Knowledge Distillation for High-Performance Mobile NPUs

arXiv: https://arxiv.org/abs/2605.03680 · HTML: https://mlsanity.com/paper/2605.03680/

## Key facts

- Cited by 2 papers in the map (data through 14 Oct 2026).
- Builds on 22 of its 29 references that are in the map.
- First posted on arXiv 5 May 2026.

## Builds on (most cited first, 22 of 22)

- [U-Net: Convolutional Networks for Biomedical Image Segmentation](https://mlsanity.com/paper/1505.04597/) (2015)
- [Distilling the Knowledge in a Neural Network](https://mlsanity.com/paper/1503.02531/) (2015)
- [SwinIR: Image Restoration Using Swin Transformer](https://mlsanity.com/paper/2108.10257/) (2021)
- [FitNets: Hints for Thin Deep Nets](https://mlsanity.com/paper/1412.6550/) (2014)
- [Restormer: Efficient Transformer for High-Resolution Image Restoration](https://mlsanity.com/paper/2111.09881/) (2021)
- [Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising](https://mlsanity.com/paper/1608.03981/) (2016)
- [Uformer: A General U-Shaped Transformer for Image Restoration](https://mlsanity.com/paper/2106.03106/) (2021)
- [Multi-Stage Progressive Image Restoration](https://mlsanity.com/paper/2102.02808/) (2021)
- [Simple Baselines for Image Restoration](https://mlsanity.com/paper/2204.04676/) (2022)
- [Learning Enriched Features for Real Image Restoration and Enhancement](https://mlsanity.com/paper/2003.06792/) (2020)
- [Toward Convolutional Blind Denoising of Real Photographs](https://mlsanity.com/paper/1807.04686/) (2018)
- [HINet: Half Instance Normalization Network for Image Restoration](https://mlsanity.com/paper/2105.06086/) (2021)
- [Benchmarking Denoising Algorithms with Real Photographs](https://mlsanity.com/paper/1707.01313/) (2017)
- [Real Image Denoising with Feature Attention](https://mlsanity.com/paper/1904.07396/) (2019)
- [Practical Deep Raw Image Denoising on Mobile Devices](https://mlsanity.com/paper/2010.06935/) (2020)
- [NTIRE 2020 Challenge on Real Image Denoising: Dataset, Methods and Results](https://mlsanity.com/paper/2005.04117/) (2020)
- [LEMUR Neural Network Dataset: Towards Seamless AutoML](https://mlsanity.com/paper/2504.10552/) (2025)
- [Fast Camera Image Denoising on Mobile GPUs with Deep Learning, Mobile AI 2021 Challenge: Report](https://mlsanity.com/paper/2105.08629/) (2021)
- [Lightweight Modules for Efficient Deep Learning based Image Restoration](https://mlsanity.com/paper/2007.05835/) (2020)
- [Multiple Degradation and Reconstruction Network for Single Image Denoising via Knowledge Distillation](https://mlsanity.com/paper/2204.13873/) (2022)
- [Lightweight network towards real-time image denoising on mobile devices](https://mlsanity.com/paper/2211.04687/) (2022)
- [LIR: A Lightweight Baseline for Image Restoration](https://mlsanity.com/paper/2402.01368/) (2024)

## Built on by (most cited first, 2 of 2)

- [Systematic Exploration of 4-Expert Heterogeneous Mixture-of-Experts via Automated Pipeline Search](https://mlsanity.com/paper/2606.23739/) (2026)
- [Towards Robust Training in NNGPT AutoML Pipeline: A Loss-Optimizer Pairing Selection Study](https://mlsanity.com/paper/2606.20933/) (2026)

## Abstract (arXiv)

While deep-learning-based image restoration has achieved unprecedented fidelity, deployment on mobile Neural Processing Units (NPUs) remains bottlenecked by operator incompatibility and memory-access overhead. We propose an NPU-aware hardware-algorithm co-design approach for real-world image denoising on mobile NPUs. Our approach employs a high-capacity teacher to supervise a lightweight student network specifically designed to leverage the tiled-memory architectures of modern mobile SoCs. By prioritizing NPU-native primitives -- standard 3x3 convolutions, ReLU activations, and nearest-neighbor upsampling -- and employing a progressive context expansion strategy (up to 1024x1024 crops), the model achieves 37.66 dB PSNR / 0.9278 SSIM on the validation benchmark and 37.58 dB PSNR / 0.9098 SSIM on the held-out test benchmark at full resolution (2432x3200) in the Mobile AI 2026 challenge. Following the official challenge rules, the inference runtime is measured under a standardized Full HD (1088x1920) protocol, where it runs in 34.0 ms on the MediaTek Dimensity 9500 and 46.1 ms on the Qualcomm Snapdragon 8 Elite NPU. We further reveal an "Inference Inversion" effect, where strict adherence to NPU-compatible operations enables dedicated NPU execution up to 3.88x faster than the integrated mobile GPU. The 1.96M-parameter student recovers 99.8% of the teacher's restoration quality via high-alpha knowledge distillation (alpha = 0.9), achieving a 21.2x parameter reduction while closing the PSNR gap from 1.63 dB to only 0.05 dB. These results establish hardware-aware distillation as an effective strategy for unifying high-fidelity denoising with practical deployment across diverse mobile NPU architectures. The proposed lightweight student model (LiteDenoiseNet) and its training statistics are provided in the NN Dataset, available at https://github.com/ABrain-One/NN-Dataset.