Fast Context-Based Low-Light Image Enhancement via Neural Implicit Representations

TL;DR

Low-light image enhancement for high-resolution imagery remains challenging due to domain gaps and computational bottlenecks. The authors propose CoLIE, a zero-shot LLIE framework that uses neural implicit representations to map 2D coordinates to the illumination field conditioned on local context in the HSV color space, estimating $\hat{x}_V = \mathbf{y}_V + f_\theta(\cdot)$ and deriving $\hat{z}_V = \mathbf{y}_V / \hat{x}_V$ under Retinex. The model employs a two-branch MLP with SIREN-style activations to fuse pixel coordinates $(p_i)$ and context $\mathcal{N}(p_i)$, and it uses a guided filter to upscale from low-resolution predictions to full resolution efficiently. Experiments on UHD-LL, MIT FiveK, and microscopy demonstrate competitive PSNR/SSIM, strong qualitative restoration, and improved downstream detection, highlighting practical impact in real-world low-light conditions.

Abstract

Current deep learning-based low-light image enhancement methods often struggle with high-resolution images, and fail to meet the practical demands of visual perception across diverse and unseen scenarios. In this paper, we introduce a novel approach termed CoLIE, which redefines the enhancement process through mapping the 2D coordinates of an underexposed image to its illumination component, conditioned on local context. We propose a reconstruction of enhanced-light images within the HSV space utilizing an implicit neural function combined with an embedded guided filter, thereby significantly reducing computational overhead. Moreover, we introduce a single image-based training loss function to enhance the model's adaptability to various scenes, further enhancing its practical applicability. Through rigorous evaluations, we analyze the properties of our proposed framework, demonstrating its superiority in both image quality and scene adaptability. Furthermore, our evaluation extends to applications in downstream tasks within low-light scenarios, underscoring the practical utility of CoLIE. The source code is available at https://github.com/ctom2/colie.

Figures (10)

• Figure 1: An example of a challenging real night image from the MIT dataset fivek that illustrates the LLIE task. The intense light emitted by the sign results in significant under-exposure of other areas in the image. Our approach excels in recovering these regions with superior accuracy compared to state-of-the-art (SOTA) methods, which tend to overexpose well-lit areas and distort colors. We emphasize that the compared methods were pretrained on challenging low-light images, whereas our method restores the image in a zero-shot setting without any prior learning.
• Figure 2: Instances of inferred illumination components represented by heatmaps based on images within the MIT dataset fivek. These estimated illumination components are subsequently used to improve low-light images according to the principles of the Retinex theory. Leveraging this approach, the enhancement process aims to faithfully reproduce the perceived visual richness and detail even in challenging lighting conditions.
• Figure 3: The motivation behind the design decisions of our method stems from a comparative analysis of the HSV components of normal-light and low-light images from the UHD-LL dataset Li2023ICLR. This overview shows that, while the Hue and Saturation components display only minimal differences, the primary distinction between the two images lies in the Value component.
• Figure 4: Our proposed framework begins with the extraction of the Value component from the HSV image representation. Subsequently, we employ a NIR model to infer the illumination component which is an essential part for effective enhancement of the input low-light image. This refined Value component is then reintegrated with the original Hue and Saturation components, forming a comprehensive representation of the enhanced image. The architecture of CoLIE involves dividing the inputs into two distinct parts: the elements of the Value component and the coordinates of the image. Each of these components is subject for regularization with unique parameters within their respective branches. By adopting this structured approach, our framework ensures precise control over the enhancement process.
• Figure 5: Visual quality comparison with SOTA methods on a real-world low-light image from the MIT dataset. Our method does not over-expose well-lit regions and preserves the original colors.