Noise Self-Regression: A New Learning Paradigm to Enhance Low-Light Images Without Task-Related Data

TL;DR

This work introduces Noise SElf-Regression (NoiSER), a data-free LLIE framework that trains a lightweight SRM using pure Gaussian noise as input and supervision. By leveraging image self-regression, a Gray-world prior, and an instance-normalization-enhanced CNN, NoiSER achieves competitive enhancement with minimal data and computation, and can automatically suppress overexposure while enabling joint-task extensions such as deraining. The paper also develops two auxiliary self-regression schemes, Pure-Color Regression (C-Regression) and Palette Regression (P-Regression), and a mean-shifted variant (MS-NoiSER) to further adapt to dataset statistics. Extensive experiments on LOL, LSRW, and SICE demonstrate strong generalization, high visual quality, and favorable application metrics compared with data-hungry baselines. Overall, NoiSER provides a lightweight, data-efficient LLIE paradigm with practical impact for robust, real-time image enhancement and potential for joint low-level vision tasks.

Abstract

Deep learning-based low-light image enhancement (LLIE) is a task of leveraging deep neural networks to enhance the image illumination while keeping the image content unchanged. From the perspective of training data, existing methods complete the LLIE task driven by one of the following three data types: paired data, unpaired data and zero-reference data. Each type of these data-driven methods has its own advantages, e.g., zero-reference data-based methods have very low requirements on training data and can meet the human needs in many scenarios. In this paper, we leverage pure Gaussian noise to complete the LLIE task, which further reduces the requirements for training data in LLIE tasks and can be used as another alternative in practical use. Specifically, we propose Noise SElf-Regression (NoiSER) without access to any task-related data, simply learns a convolutional neural network equipped with an instance-normalization layer by taking a random noise image, $\mathcal{N}(0,σ^2)$ for each pixel, as both input and output for each training pair, and then the low-light image is fed to the trained network for predicting the normal-light image. Technically, an intuitive explanation for its effectiveness is as follows: 1) the self-regression reconstructs the contrast between adjacent pixels of the input image, 2) the instance-normalization layer may naturally remediate the overall magnitude/lighting of the input image, and 3) the $\mathcal{N}(0,σ^2)$ assumption for each pixel enforces the output image to follow the well-known gray-world hypothesis when the image size is big enough. Compared to current state-of-the-art LLIE methods with access to different task-related data, NoiSER is highly competitive in enhancement quality, yet with a much smaller model size, and much lower training and inference cost. Besides, NoiSER also excels in mitigating overexposure and handling joint tasks.

Figures (18)

• Figure 1: Comparison of training data settings used in current LLIE methods, including the paired data in RetinexNet RetinexNet, unpaired data in EnlightenGAN EnlightenGAN, zero-reference data in Zero-DCE Zero-DCE, and the noise (task-irrelevant data) in our proposed NoiSER. Clearly, NoiSER performs the best in naturalness and detail recovery.
• Figure 2: The training and inference pipeline of NoiSER. During training, NoiSER just samples noise $n\sim \mathcal{N}(0,\sigma^2)$ as both model input and supervised signal to train a self-regression model (SRM), i.e., $n^\ast$$=$$SRM(n)$$\approx$$n$. During inference, the trained SRM can directly enhance low-light images. Note that the designed SRM is equipped with one instance normalization and two non-linear activation layers without shortcuts, which avoids learning an identity mapping.
• Figure 3: Comparison of the fully-converged enhancement results using different training data. Clearly, the noise self-regression approach can yield visually better results.
• Figure 4: Iteration-Distance curves using different pure-colors for C-regression training. The distance is the $L_2$ norm between the output and the pure-central-grey during training. The pure-central-grey curve did not converge to $0$ but to a very small number.
• Figure 5: The phenomenon obtained from the pure-black self-regression. For an arbitrary input with a degree of contrast, the model finally tend to divide the input to be either black or white.