DarkDeblur: Learning single-shot image deblurring in low-light condition

TL;DR

This work tackles single-shot image deblurring in low-light environments by introducing DarkDeblurNet, a GAN-enabled generator that employs a dense-attention block and contextual gating within a feature pyramid to capture global and spatial dependencies. It optimizes a multi-term loss comprising reconstruction, structural, perceptual, and adversarial components, enabling visually plausible restorations under challenging lighting. The approach is validated on synthesized (ExDark, Lai, Kohler) and real-world (DarkShake) datasets and demonstrates superior quantitative and qualitative results, as well as improvements in downstream tasks like segmentation, detection, and OCR. The paper also contributes the DarkShake real-world dataset and discusses practical considerations, limitations, and future extensions such as unsupervised training and multi-frame deblurring.

Abstract

Single-shot image deblurring in a low-light condition is known to be a profoundly challenging image translation task. This study tackles the limitations of the low-light image deblurring with a learning-based approach and proposes a novel deep network named as DarkDeblurNet. The proposed DarkDeblur- Net comprises a dense-attention block and a contextual gating mechanism in a feature pyramid structure to leverage content awareness. The model additionally incorporates a multi-term objective function to perceive a plausible perceptual image quality while performing image deblurring in the low-light settings. The practicability of the proposed model has been verified by fusing it in numerous computer vision applications. Apart from that, this study introduces a benchmark dataset collected with actual hardware to assess the low-light image deblurring methods in a real-world setup. The experimental results illustrate that the proposed method can outperform the state-of-the-art methods in both synthesized and real-world data for single-shot image deblurring, even in challenging lighting environment.

Figures (16)

• Figure 1: Performance of the SOTA single-shot deblurring methods and proposed DarkDeblurNet on a real-world blurry image. It is visible that existing methods likely to produce visually disturbing artifacts while removing blurs from an image capture in low-light. (a) Blurry input image (PSNR: 23.7). (b) Result of LightStreaks (PSNR: 15.73 ). (c) Result of DarkChannelpan2016blind (PSNR:23.44). (d) Result of DeepDeblur nah2017deep (PSNR: 25.89). (e) Result of SRN tao2018scale (PSNR: 25.77). (f) Result of DeblurGANv2 kupyn2019deblurgan (PSNR: 24.44). (g) DarkDeblurNet (PSNR:27.99). (h) Reference sharp image.
• Figure 2: The overview of the proposed DarkDeblurNet. The proposed network incorporates a novel dense-attention block and contextual gating mechanism in a feature pyramid structure. Also, it follows the principle of generative adversarial networks. (a) The architecture of the generator. (b) The architecture of the discriminator.
• Figure 3: Overview of the proposed dense-attention block. The dense-attention block combines a residual dense block and channel attention mechanism. It aims to capture the global feature interdependencies in different feature levels.
• Figure 4: Overview of the contextual gate. The contextual-gate aims to propagate the spatial dependencies between different feature levels.
• Figure 5: Example of a blur-sharp image pair obtained from the ExDark dataset. (a) Reference (sharp) image. (b) Blurry (simulated) image.