Why Low-Light Cameras Go Color Blind: Removing Color Bias in Raw Denoising

Low-light raw image denoising is critical for high-quality photography, yet existing methods often struggle with color shifts and chromatic artifacts in dark regions. The authors investigate why these artifacts occur, focusing on the role of black-level error (BLE)—the discrepancy between a sensor's recorded black-level metadata and its actual physical state—and seek a camera-agnostic, calibration-free solution to this problem.

The researchers analyze how different noise formulations contribute to color bias. They find that standard denoising pipelines often fail because they do not explicitly account for BLE, which is amplified significantly during the high-gain processing required for low-light images. To address this, they introduce a Black-Level Bias Estimator (BLBE) network. This module predicts the BLE as a global feature of the noisy input, allowing the system to subtract the error before the image is passed to a generic, camera-agnostic denoising model. They also identify that the widely used SIDD dataset contains ground-truth color bias and release a corrected version, SIDD-CC, to provide a more reliable benchmark.

The study demonstrates that incorporating a learnable BLE correction step significantly improves color reproduction in low-light denoising. Their approach achieves state-of-the-art performance among blind (calibration-free) denoisers and is often competitive with, or superior to, methods that require expensive camera-specific calibration or large-scale paired training data. Furthermore, the authors show that even some supervised, camera-specific models suffer from color bias because their ground-truth training data itself contains distortions, which the new SIDD-CC benchmark helps to mitigate.

This work provides a practical, scalable solution for low-light denoising that avoids the labor-intensive process of per-device calibration or collecting massive amounts of paired training data. By identifying BLE as a fundamental source of color degradation, the authors offer a more robust framework for computational photography that generalizes across different camera sensors, while also improving the quality of standard benchmarks used by the research community.