Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Latent Disentanglement for Low Light Image Enhancement

Zhihao Zheng, Mooi Choo Chuah

TL;DR

This work introduces LDE-Net, a transformer-based latent-disentanglement framework for low-light image enhancement that separates content and illumination in latent space to avoid corruption from pixel-space Retinex decompositions. A Content-Aware Embedding (CAE) module guides a light-weight illumination enhancer, producing strong LLIE results and enabling efficient downstream tasks such as nighttime UAV tracking and low-light object detection. The approach achieves state-of-the-art performance on public LLIE benchmarks (LOL-v1, LOL-v2, LDIS) and demonstrates transferability and practicality through ablations and downstream task experiments. By combining robust latent disentanglement with cross-attention-based fusion, the method delivers sharper details, better lighting consistency, and competitive or superior performance with significantly reduced computational cost.

Abstract

Many learning-based low-light image enhancement (LLIE) algorithms are based on the Retinex theory. However, the Retinex-based decomposition techniques in such models introduce corruptions which limit their enhancement performance. In this paper, we propose a Latent Disentangle-based Enhancement Network (LDE-Net) for low light vision tasks. The latent disentanglement module disentangles the input image in latent space such that no corruption remains in the disentangled Content and Illumination components. For LLIE task, we design a Content-Aware Embedding (CAE) module that utilizes Content features to direct the enhancement of the Illumination component. For downstream tasks (e.g. nighttime UAV tracking and low-light object detection), we develop an effective light-weight enhancer based on the latent disentanglement framework. Comprehensive quantitative and qualitative experiments demonstrate that our LDE-Net significantly outperforms state-of-the-art methods on various LLIE benchmarks. In addition, the great results obtained by applying our framework on the downstream tasks also demonstrate the usefulness of our latent disentanglement design.

Latent Disentanglement for Low Light Image Enhancement

TL;DR

This work introduces LDE-Net, a transformer-based latent-disentanglement framework for low-light image enhancement that separates content and illumination in latent space to avoid corruption from pixel-space Retinex decompositions. A Content-Aware Embedding (CAE) module guides a light-weight illumination enhancer, producing strong LLIE results and enabling efficient downstream tasks such as nighttime UAV tracking and low-light object detection. The approach achieves state-of-the-art performance on public LLIE benchmarks (LOL-v1, LOL-v2, LDIS) and demonstrates transferability and practicality through ablations and downstream task experiments. By combining robust latent disentanglement with cross-attention-based fusion, the method delivers sharper details, better lighting consistency, and competitive or superior performance with significantly reduced computational cost.

Abstract

Many learning-based low-light image enhancement (LLIE) algorithms are based on the Retinex theory. However, the Retinex-based decomposition techniques in such models introduce corruptions which limit their enhancement performance. In this paper, we propose a Latent Disentangle-based Enhancement Network (LDE-Net) for low light vision tasks. The latent disentanglement module disentangles the input image in latent space such that no corruption remains in the disentangled Content and Illumination components. For LLIE task, we design a Content-Aware Embedding (CAE) module that utilizes Content features to direct the enhancement of the Illumination component. For downstream tasks (e.g. nighttime UAV tracking and low-light object detection), we develop an effective light-weight enhancer based on the latent disentanglement framework. Comprehensive quantitative and qualitative experiments demonstrate that our LDE-Net significantly outperforms state-of-the-art methods on various LLIE benchmarks. In addition, the great results obtained by applying our framework on the downstream tasks also demonstrate the usefulness of our latent disentanglement design.
Paper Structure (17 sections, 11 equations, 5 figures, 5 tables)

This paper contains 17 sections, 11 equations, 5 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Traditional LLIE methods
  4. Learning-based LLIE methods
  5. Methodologies
  6. Motivation
  7. Overview
  8. Latent Disentanglement
  9. Content-Aware Illumination Enhancement
  10. Loss functions for LLIE
  11. Experiments
  12. Experimental Settings
  13. Low-light Image Enhancement
  14. Nighttime UAV Tracking
  15. Low-light Object Detection
  16. ...and 2 more sections

Figures (5)

  • Figure 1: Overview of the proposed LDE-Net. (a)The disentanglement module disentangle an input image to the Content and Illumination components, both of which are latent space features. (b)Content-Aware Illumination Enhancement module enhances the Illumination with the guidance of Content feature. (c)The reconstruction module reconstructs a new image with the Content and restored Illumination component.
  • Figure 2: Illustration of the implementation details of the Disentanglement Transformer Block. (DTB)
  • Figure 3: Illustration of the implementation details of the Content-Aware Embedding Module (CAE).
  • Figure 4: Qualitative results of LOL (top) and LDIS (bottom). Previous methods either collapse by noise, or distort color, or produce blurry and under-/over-exposed images. While our algorithm can effectively reconstruct well-exposed image details.
  • Figure 5: Overview of the light-weight enhancer with latent disentanglement module.