Low-Light Image Enhancement with Wavelet-based Diffusion Models (2306.00306v3)

Abstract: Diffusion models have achieved promising results in image restoration tasks, yet suffer from time-consuming, excessive computational resource consumption, and unstable restoration. To address these issues, we propose a robust and efficient Diffusion-based Low-Light image enhancement approach, dubbed DiffLL. Specifically, we present a wavelet-based conditional diffusion model (WCDM) that leverages the generative power of diffusion models to produce results with satisfactory perceptual fidelity. Additionally, it also takes advantage of the strengths of wavelet transformation to greatly accelerate inference and reduce computational resource usage without sacrificing information. To avoid chaotic content and diversity, we perform both forward diffusion and denoising in the training phase of WCDM, enabling the model to achieve stable denoising and reduce randomness during inference. Moreover, we further design a high-frequency restoration module (HFRM) that utilizes the vertical and horizontal details of the image to complement the diagonal information for better fine-grained restoration. Extensive experiments on publicly available real-world benchmarks demonstrate that our method outperforms the existing state-of-the-art methods both quantitatively and visually, and it achieves remarkable improvements in efficiency compared to previous diffusion-based methods. In addition, we empirically show that the application for low-light face detection also reveals the latent practical values of our method. Code is available at https://github.com/JianghaiSCU/Diffusion-Low-Light.

Low-Light Image Enhancement with Wavelet-based Diffusion Models: An Evaluation

The paper "Low-Light Image Enhancement with Wavelet-based Diffusion Models" addresses significant challenges in the domain of computational photography, specifically focusing on enhancing images captured in low-light conditions. This is a pivotal research contribution in image restoration tasks, leveraging diffusion models, which have been gaining traction due to their capability in high-quality image synthesis.

The authors propose a novel approach termed DiffLL, which integrates wavelet transformation within the diffusion model framework to improve performance efficiency and achieve high-fidelity image restoration. This approach employs a Wavelet-based Conditional Diffusion Model (WCDM) that exploits wavelet transformation to condense information spatially without loss, enabling reduced computational demand and accelerated inference. This is contrasted with traditional methods where operations are typically conducted in either image space or latent space, which are computationally expensive.

A notable aspect of this work is the dual-process training strategy that incorporates both forward diffusion and denoising during training, enhancing model stability and minimizing the randomness of inference outputs. This training approach distinguishes itself by its ability to generate consistent, high-quality outputs without content chaos often introduced by stochastic sampling processes inherent to standard diffusion models.

The implementation of a High-Frequency Restoration Module (HFRM) reflects the model's sophistication, utilizing cross-attention mechanisms to integrate vertical, horizontal, and diagonal information. This integration is crucial for fine-grained detail reconstruction and addresses the often-overlooked aspect of detail restoration in low-light image enhancement tasks.

Quantitative evaluations on several datasets, including LOLv1, LOLv2-real, and LSRW, illustrate the superior performance of the authors' method compared to existing state-of-the-art methods. On distortion metrics such as PSNR and SSIM, DiffLL consistently outperformed alternatives, validating its effectiveness. The model's perceptual quality as measured by LPIPS and FID also demonstrates reduced artifacts and improved aesthetic qualities, proving its applicability in realistic scenarios.

The reduction in computational overhead is another highlight of this research, with DiffLL showing a notable increase in efficiency, being up to 70 times faster than comparable models like DDIM while sustaining comparable or superior qualitative performance. This is critical for practical deployment in computationally constrained environments or real-time applications.

The authors also explore potential applications in low-light face detection, showing improved performance of face detectors when pre-processing images with their method. This highlights the broader applicability and potential for integration of DiffLL in diverse downstream tasks.

Speculation on future advancements sees this method being adapted for higher resolution image enhancement and possibly extended to a variety of other challenging image restoration tasks. The efficient training methodology coupled with robust restoration quality positions DiffLL as a significant step forward in low-light image enhancement.

In conclusion, this research effectively addresses key deficiencies in existing diffusion-based enhancement methods by integrating wavelet transformations for resource-efficient, high-quality image restoration. The documented numerical superiority on extensive benchmarks and its versatility in enhancing lower-light face detection verifies its comprehensive value to both theoretical research and practical applications within computational photography. Future work could explore the application of this framework across different image restoration tasks to further validate its robustness and utility in various real-world conditions.