Feature-Align Network with Knowledge Distillation for Efficient Denoising

We propose an efficient neural network for RAW image denoising. Although neural network-based denoising has been extensively studied for image restoration, little attention has been given to efficient denoising for compute limited and power sensitive devices, such as smartphones and smartwatches. In this paper, we present a novel architecture and a suite of training techniques for high quality denoising in mobile devices. Our work is distinguished by three main contributions. (1) Feature-Align layer that modulates the activations of an encoder-decoder architecture with the input noisy images. The auto modulation layer enforces attention to spatially varying noise that tend to be "washed away" by successive application of convolutions and non-linearity. (2) A novel Feature Matching Loss that allows knowledge distillation from large denoising networks in the form of a perceptual content loss. (3) Empirical analysis of our efficient model trained to specialize on different noise subranges. This opens additional avenue for model size reduction by sacrificing memory for compute. Extensive experimental validation shows that our efficient model produces high quality denoising results that compete with state-of-the-art large networks, while using significantly fewer parameters and MACs. On the Darmstadt Noise Dataset benchmark, we achieve a PSNR of 48.28dB, while using 263 times fewer MACs, and 17.6 times fewer parameters than the state-of-the-art network, which achieves 49.12dB.

Feature-Align Network and Knowledge Distillation for Efficient Denoising

Deep learning-based RAW image denoising is a quintessential problem in image restoration. Recent works have pushed the state-of-the-art in denoising image quality. However, many of these networks are computationally too expensive for efficient use in mobile devices. Here, we propose a novel network for efficient RAW denoising on mobile devices. Our contributions are: (1) An efficient encoder-decoder network augmented with a new Feature-Align layer to attend to spatially varying noise. (2) A new perceptual Feature Loss calculated in the RAW domain to preserve high frequency image content. (3) An analysis of the use of multiple models tuned to different subranges of noise levels. (4) An open-source RAW noisy-clean paired dataset with noise modeling, to facilitate research in RAW denoising. We evaluate the effectiveness of our proposed network and training techniques and show results that compete with the state-of-the-art network, while using significantly fewer parameters and MACs. On the Darmstadt Noise Dataset benchmark, we achieve a PSNR of 48.28dB, with 263 times fewer MACs, and 17.6 times fewer parameters than the state-of-the-art network, which achieves 49.12 dB.