Attention Guided Low-light Image Enhancement with a Large Scale Low-light Simulation Dataset (1908.00682v3)

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.

• The paper introduces an attention-guided dual-map CNN that adaptively enhances brightness and reduces noise in low-light images.
• It leverages a large-scale synthetic low-light simulation dataset to train more robust models across diverse conditions.
• Extensive experiments show improved fidelity and reduced artifacts, highlighting its potential in surveillance and autonomous driving.

Attention Guided Low-light Image Enhancement with a Large Scale Low-light Simulation Dataset

The paper presents an advanced method for enhancing low-light images, focusing on the challenges associated with noise, color distortion, and brightness recovery. The proposed technique leverages an attention-guided multi-branch convolutional neural network (CNN) architecture, complemented by a newly constructed large-scale low-light simulation dataset. This synthetic dataset is tailored to provide a comprehensive range of conditions and facilitates the training of more robust models.

Methodology Overview

The core innovation of the paper is an end-to-end attention-guided system that utilizes dual attention maps: one for brightness enhancement and another for noise reduction. These attention maps are pivotal as they direct the enhancement network to focus on specific regions of the image: the underexposed regions for brightness correction and the areas containing noise for denoising. This is attained through a process of multi-branch decomposition and fusion, which allows the network to enhance images adaptively based on input characteristics.

1. Synthetic Dataset Construction: The authors propose a synthetic dataset with a significantly larger scale and diversity compared to existing datasets. This dataset is created using carefully crafted low-light simulation strategies, which include unique augmentation techniques to simulate realistic low-light conditions.
2. Attention Maps: The methodology employs two distinct attention maps. The first distinguishes between underexposed and well-lit regions, guiding brightness enhancement. The second map differentiates noise from textures, thereby aiding in accurate denoising.
3. Network Architecture: The enhancement process is handled by a multi-branch CNN where each branch is tasked with processing different elements of the image. The core network comprises feature extraction, enhancement, and fusion modules. The Reinforcement-Net component further refines the color and contrast of enhanced images.

Experimental Results

The paper conducts extensive experiments on multiple datasets, illustrating significant improvements over existing state-of-the-art methods. Quantitatively and visually, the proposed method shows superior performance in fidelity and enhancement quality. Importantly, the results underscore the effectiveness of the dual attention mechanism in producing high-quality enhancements without introducing artifacts commonly associated with standard brightness adjustment techniques.

Implications and Future Directions

The research presents substantial practical implications, especially in scenarios like autonomous driving and surveillance, where image quality in low-light conditions is critical. Moreover, the novel dataset provides a valuable resource for future low-light image enhancement studies. Theoretically, the work advances our understanding of attention-guided mechanisms in complex image processing tasks.

Looking forward, potential advancements could explore the integration of this approach with video data to address low-light enhancement across temporal sequences. Moreover, further investigation into alternative architectures or learning paradigms, such as self-supervised learning, could offer additional improvements and applications across a broader range of imaging conditions and modalities.

This paper represents a notable contribution to the field of image enhancement by combining innovative neural network architectures with a robust data-centric approach, paving the way for better-performing models across diverse low-light scenarios.