Edge-guided Low-light Image Enhancement with Inertial Bregman Alternating Linearized Minimization

Chaoyan Huang, Zhongming Wu, Tieyong ZengarXiv 2024 · DOI · arXiv

Abstract

Low-light image enhancement is a critical task for computer vision applications, yet it remains challenging due to low contrast and heavy noise. This paper proposes a hybrid framework that combines the feature-extraction capabilities of deep learning with the mathematical interpretability of variational Retinex models. We introduce an "Edge-Net" to extract structural priors from dark images and a novel Inertial Bregman Alternating Linearized Minimization (IBALM) algorithm to solve the resulting optimization problem. Our approach provides both state-of-the-art performance and rigorous theoretical convergence guarantees.

Introduction

Images captured in low-light environments often suffer from degraded visibility, lost structural details, and significant noise. While traditional Retinex-based methods attempt to decompose images into illumination and reflectance, they frequently fail because standard edge detection is unreliable in high-noise, low-contrast conditions. Conversely, pure deep learning methods, while effective, often function as "black boxes" lacking theoretical stability.

This research addresses these issues by learning a robust edge prior via a dedicated neural network and integrating it into a mathematically sound variational framework. The core innovation lies in the IBALM algorithm, which ensures efficient optimization and guaranteed convergence to a stationary point.

Proposed Methodology

The framework is divided into two primary stages: edge prior learning and variational optimization.

Edge-Net Architecture

To overcome the limitations of traditional edge detectors in the dark, we employ Edge-Net, a convolutional neural network based on Multi-Scale Residual Blocks (MSRB). Trained on the LOL (Low-Light) dataset, Edge-Net is designed to identify clean structural edges directly from degraded inputs. These learned edges serve as a crucial regularizer for the subsequent enhancement step.

Variational Retinex Model and IBALM

We formulate the enhancement problem in the logarithmic domain to linearize the Retinex decomposition ($S = R \circ L$). The optimization problem is solved using the Inertial Bregman Alternating Linearized Minimization (IBALM) algorithm.

Overview of the proposed framework, including the Edge-Net architecture and the IBALM optimization process.

IBALM incorporates two key features:

  1. Inertial Terms: A momentum-based technique that utilizes information from previous iterations to accelerate convergence.
  2. Bregman Distance: This replaces standard Euclidean distance, allowing the algorithm to handle non-smooth regularizers more effectively.

Unlike many hybrid models, the authors provide a mathematical proof of convergence using the Kurdyka-Łojasiewicz property, ensuring the algorithm's stability.

Experimental Results

The proposed method was evaluated against several state-of-the-art (SOTA) models using both paired datasets (e.g., FiveK, LSRW) and real-world non-reference datasets (e.g., LIME, DICM).

Quantitative and Qualitative Performance

On the FiveK dataset, the method achieved a PSNR of 17.82 dB, significantly outperforming the SMG model's 16.31 dB. Qualitative analysis reveals that the approach preserves sharper edges and avoids the over-enhancement or "washed-out" appearance common in competing Retinex methods.

Visual comparison of enhancement results and edge extraction quality between the proposed method, SMG, and ground truth.

As shown in the visual results, the learned edge prior (g) closely matches the ground truth (h), whereas edges extracted directly from the input (e) or other models (f) are either noisy or overly blurred.

Conclusion

By combining deep learning for prior extraction with a rigorous optimization framework, this work sets a new standard for low-light image enhancement. The IBALM algorithm bridges the gap between performance and interpretability. Future work will focus on addressing color constancy and non-uniform lighting artifacts in extreme real-world scenarios.

Summary updated

Abstract

Prior-based methods for low-light image enhancement often face challenges in extracting available prior information from dim images. To overcome this limitation, we introduce a simple yet effective Retinex model with the proposed edge extraction prior. More specifically, we design an edge extraction network to capture the fine edge features from the low-light image directly. Building upon the Retinex theory, we decompose the low-light image into its illumination and reflectance components and introduce an edge-guided Retinex model for enhancing low-light images. To solve the proposed model, we propose a novel inertial Bregman alternating linearized minimization algorithm. This algorithm addresses the optimization problem associated with the edge-guided Retinex model, enabling effective enhancement of low-light images. Through rigorous theoretical analysis, we establish the convergence properties of the algorithm. Besides, we prove that the proposed algorithm converges to a stationary point of the problem through nonconvex optimization theory. Furthermore, extensive experiments are conducted on multiple real-world low-light image datasets to demonstrate the efficiency and superiority of the proposed scheme.
BibTeX
@misc{huang2024edge,
  author = {Chaoyan Huang and Zhongming Wu and Tieyong Zeng},
  title = {{Edge-guided Low-light Image Enhancement with Inertial Bregman Alternating Linearized Minimization}},
  year = {2024},
  eprint = {2403.01142},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2403.01142},
}
Figure