[Paper Review] Pseudo-ISP: Learning Pseudo In-camera Signal Processing Pipeline from A Color Image Denoiser

Yue Cao, Xiaohe Wu+4 more|arXiv (Cornell University)|Mar 18, 2021

Image and Signal Denoising Methods46 references4 citations

TL;DR

This paper proposes Pseudo-ISP, an unpaired learning framework that adapts pre-trained color image denoisers to real-world noisy images with noise discrepancy by jointly learning a pseudo in-camera signal processing (ISP) pipeline and signal-dependent rawRGB noise model from unpaired test noisy and clean images. By synthesizing realistic noisy sRGB images using the learned pseudo-ISP and noise model, the method enables effective denoiser fine-tuning, significantly boosting performance on diverse real-world datasets, including state-of-the-art models like CBDNet and RIDNet.

ABSTRACT

The success of deep denoisers on real-world color photographs usually relies on the modeling of sensor noise and in-camera signal processing (ISP) pipeline. Performance drop will inevitably happen when the sensor and ISP pipeline of test images are different from those for training the deep denoisers (i.e., noise discrepancy). In this paper, we present an unpaired learning scheme to adapt a color image denoiser for handling test images with noise discrepancy. We consider a practical training setting, i.e., a pre-trained denoiser, a set of test noisy images, and an unpaired set of clean images. To begin with, the pre-trained denoiser is used to generate the pseudo clean images for the test images. Pseudo-ISP is then suggested to jointly learn the pseudo ISP pipeline and signal-dependent rawRGB noise model using the pairs of test and pseudo clean images. We further apply the learned pseudo ISP and rawRGB noise model to clean color images to synthesize realistic noisy images for denoiser adaption. Pseudo-ISP is effective in synthesizing realistic noisy sRGB images, and improved denoising performance can be achieved by alternating between Pseudo-ISP training and denoiser adaption. Experiments show that our Pseudo-ISP not only can boost simple Gaussian blurring-based denoiser to achieve competitive performance against CBDNet, but also is effective in improving state-of-the-art deep denoisers, e.g., CBDNet and RIDNet.

Motivation & Objective

• To address performance degradation in deep denoisers when applied to test images with different sensor noise and ISP pipelines than training data (noise discrepancy).
• To enable adaptation of a pre-trained denoiser using only a set of test noisy images and an unpaired set of clean images, without requiring paired noisy-clean data for the test domain.
• To learn a realistic noise model and pseudo-ISP pipeline for sRGB images by leveraging a pre-trained denoiser to generate pseudo-clean images.
• To improve denoising performance through alternating training between pseudo-ISP learning and denoiser fine-tuning using synthetic noisy images.

Proposed method

• Use a pre-trained denoiser to generate pseudo-clean images from test noisy images, forming unpaired test-noisy and pseudo-clean image pairs.
• Propose a Pseudo-ISP model with three subnetworks: sRGB2Raw (inverse ISP), noise estimation (modeling signal-dependent noise in pseudo rawRGB), and Raw2sRGB (forward ISP) to simulate the camera pipeline.
• Train the Pseudo-ISP model end-to-end on the unpaired image pairs to learn a realistic rawRGB noise model and pseudo-ISP transformation.
• Use the learned Pseudo-ISP and noise model to synthesize realistic noisy sRGB images from clean images, creating synthetic paired data for denoiser fine-tuning.
• Alternate between Pseudo-ISP training and denoiser fine-tuning to iteratively improve both noise modeling and denoising performance.
• Apply the method to both weak (e.g., Gaussian blurring) and strong (e.g., CBDNet, RIDNet) pre-trained denoisers to demonstrate broad applicability.

Experimental results

Research questions

• RQ1 Can a pre-trained denoiser be effectively adapted to real-world noisy images with unknown and mismatched sensor noise and ISP pipelines using only unpaired noisy and clean images?
• RQ2 How can a realistic signal-dependent noise model and pseudo-ISP pipeline be learned from unpaired test noisy and pseudo-clean image pairs?
• RQ3 To what extent can the synthesized noisy images from the Pseudo-ISP model improve denoiser generalization and performance on unseen test datasets?
• RQ4 Does the proposed unpaired learning scheme outperform standard fine-tuning or data augmentation baselines in handling noise discrepancy?
• RQ5 Can Pseudo-ISP improve both classical and state-of-the-art deep denoisers across diverse real-world datasets with varying noise characteristics?

Key findings

• Pseudo-ISP improves the performance of a simple Gaussian blurring-based denoiser to achieve competitive results against CBDNet on real-world datasets.
• The method boosts state-of-the-art denoisers like CBDNet, RIDNet, and PT-MWRN, with PSNR gains of up to 0.53 dB (CBDNet) and 0.17 dB (RIDNet) on the DND dataset.
• On SIDDPlus, which has large noise discrepancy from the SIDD training set, Pseudo-ISP achieves a significant PSNR gain of over 0.5 dB, demonstrating strong generalization.
• Even on challenging datasets like CC15 and MIT-IP8 with JPEG compression and complex demosaicking, Pseudo-ISP achieves PSNR improvements of 0.3–0.4 dB.
• The method outperforms standard fine-tuning and data augmentation baselines by a clear margin, with gains not attributable to longer training time.
• PT-MWRN* fine-tuned with Pseudo-ISP achieves the best PSNR on all five datasets, outperforming the second-best method DIDN by 0.44 dB on SIDDPlus.