TSDN: Two-Stage Raw Denoising in the Dark.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Deep learning excels at image denoising but struggles with severe noise in low-light conditions.
• Existing state-of-the-art (SOTA) algorithms are computationally intensive, hindering real-time, high-resolution image processing and hardware compatibility.

Purpose of the Study:

• To propose a novel, lightweight, and hardware-friendly low-light RAW denoising algorithm named Two-Stage-Denoising (TSDN).
• To address the performance limitations of current denoising methods in extremely dark environments.

Main Methods:

• TSDN employs a two-stage approach: initial noise removal to create an intermediate image, followed by image restoration.
• A novel training strategy, Expand-Shrink-Learning (ESL), is introduced, involving expanding the network, then shrinking it via Channel-Shrink-Learning (CSL) and Layer-Shrink-Learning (LSL).

Main Results:

• TSDN demonstrates superior denoising performance (PSNR and SSIM) compared to SOTA algorithms in dark environments.
• The TSDN model size is significantly smaller (one-eighth) than traditional U-Net architectures, indicating improved efficiency.

Conclusions:

• TSDN effectively denoises low-light RAW images, outperforming existing methods.
• The proposed ESL training method enables efficient training of lightweight networks for real-time, hardware-friendly denoising applications.