Re-parameterized Multi-scale Fusion Network for Efficient Extreme Low-light Raw Denoising

Abstract

Practical low-light denoising/enhancement solutions often require fast computation,high memory efficiency,and can achieve visually high-quality restoration results.Most existing methods aim to restore quality but compromise on speed and memory requirements,which limits their usefulness to a large extent.This paper proposes a new deep denoising architecture,a re-parameterized multi-scale fusion network for extreme low-light raw denoising,which greatly improves the inference speed without losing high-quality denoising performance.Specifically,image features are extracted in multi-scale space,and a lightweight spatial-channel parallel attention module is used to focus on core features within space and channel dynamically and adaptively.The representation ability of the model is further enriched by re-parameterized convolutional unit without increasing computational cost at inference.The proposed model can restore UHD 4K resolution images within about 1s on a CPU(e.g.,Intel i7-7700K) and run at 24 fps on a GPU(e.g.,NVIDIA GTX 1080Ti),which is almost four times faster than existing advanced methods(e.g.,UNet) while still maintaining competitive restoration quality.

Keywords

• re-parameterization convolutional unit|multi-scale fusion|spatial-channel parallel attention module|extreme low-light denoising