[Paper Review] Attention-guided Low-light Image Enhancement.

Feifan Lv, Feng Lu|arXiv (Cornell University)|Aug 2, 2019

Image Enhancement Techniques2 references22 citations

TL;DR

This paper proposes an end-to-end attention-guided low-light image enhancement method using a multi-branch convolutional neural network. It introduces a large, diverse synthetic dataset and employs two task-specific attention maps—one for brightness enhancement and another for noise suppression—enabling adaptive, high-fidelity enhancement that outperforms state-of-the-art methods both quantitatively and visually.

ABSTRACT

Low-light image enhancement is challenging in that it needs to consider not only brightness recovery but also complex issues like color distortion and noise, which usually hide in the dark. Simply adjusting the brightness of a low-light image will inevitably amplify those artifacts. To address this difficult problem, this paper proposes a novel end-to-end attention-guided method based on multi-branch convolutional neural network. To this end, we first construct a synthetic dataset with carefully designed low-light simulation strategies. The dataset is much larger and more diverse than existing ones. With the new dataset for training, our method learns two attention maps to guide the brightness enhancement and denoising tasks respectively. The first attention map distinguishes underexposed regions from well lit regions, and the second attention map distinguishes noises from real textures. With their guidance, the proposed multi-branch decomposition-and-fusion enhancement network works in an input adaptive way. Moreover, a reinforcement-net further enhances color and contrast of the output image. Extensive experiments on multiple datasets demonstrate that our method can produce high fidelity enhancement results for low-light images and outperforms the current state-of-the-art methods by a large margin both quantitatively and visually.

Motivation & Objective

• Address the challenge of low-light image enhancement, which involves balancing brightness recovery, color fidelity, and noise suppression.
• Overcome limitations of traditional methods that amplify artifacts when simply increasing brightness.
• Develop a data-driven approach capable of distinguishing between underexposed regions, real textures, and noise in low-light conditions.
• Improve enhancement quality by learning input-adaptive attention maps for brightness and denoising tasks.
• Achieve superior performance over existing state-of-the-art methods through joint optimization of enhancement and color/contrast refinement.

Proposed method

• Construct a large-scale, diverse synthetic low-light dataset using carefully designed simulation strategies to better represent real-world low-light conditions.
• Design a multi-branch convolutional neural network that processes brightness enhancement and denoising tasks in parallel, guided by two distinct attention maps.
• Learn the first attention map to differentiate underexposed regions from well-lit areas, enabling targeted brightness recovery.
• Learn the second attention map to distinguish noise from real textures, allowing selective denoising without blurring fine details.
• Integrate a reinforcement network to further enhance color saturation and contrast in the final output image.
• Train the entire network end-to-end using a combination of perceptual, reconstruction, and adversarial losses to ensure high-fidelity results.

Experimental results

Research questions

• RQ1 Can a synthetic dataset with diverse low-light conditions improve the generalization and performance of low-light image enhancement networks?
• RQ2 Can task-specific attention maps effectively guide brightness enhancement and noise suppression in a single unified framework?
• RQ3 Does input-adaptive processing via learned attention maps lead to better preservation of textures and reduction of artifacts compared to non-adaptive methods?
• RQ4 Can the integration of a reinforcement network further improve color and contrast quality in low-light enhanced images?
• RQ5 How does the proposed method compare quantitatively and qualitatively to state-of-the-art low-light enhancement approaches?

Key findings

• The proposed method achieves superior quantitative performance on multiple benchmark datasets, outperforming existing state-of-the-art methods in PSNR, SSIM, and LPIPS metrics.
• Visual results demonstrate significant improvements in brightness recovery, color accuracy, and noise suppression without introducing artifacts or over-enhancing textures.
• The use of a large, diverse synthetic dataset leads to better generalization across real-world low-light scenarios compared to methods trained on smaller or less varied data.
• The two-task attention mechanism effectively separates and addresses distinct enhancement challenges—brightness and denoising—leading to more precise and adaptive processing.
• The reinforcement network successfully enhances color and contrast, contributing to more natural-looking results that are preferred in user studies.
• Ablation studies confirm that both attention maps and the reinforcement network contribute significantly to the final performance gain.