Abstract:Diffusion model-based low-light image enhancement methods rely heavily on paired training data, leading to limited extensive application. Meanwhile, existing unsupervised methods lack effective bridging capabilities for unknown degradation. To address these limitations, we propose a novel zero-reference lighting estimation diffusion model for low-light image enhancement called Zero-LED. It utilizes the stable convergence ability of diffusion models to bridge the gap between low-light domains and real normal-light domains and successfully alleviates the dependence on pairwise training data via zero-reference learning. Specifically, we first design the initial optimization network to preprocess the input image and implement bidirectional constraints between the diffusion model and the initial optimization network through multiple objective functions. Subsequently, the degradation factors of the real-world scene are optimized iteratively to achieve effective light enhancement. In addition, we explore a frequency-domain based and semantically guided appearance reconstruction module that encourages feature alignment of the recovered image at a fine-grained level and satisfies subjective expectations. Finally, extensive experiments demonstrate the superiority of our approach to other state-of-the-art methods and more significant generalization capabilities. We will open the source code upon acceptance of the paper.
What problem does this paper attempt to address?
The problems that this paper attempts to solve are several key challenges in low - light image enhancement techniques:
1. **Dependence on paired training data**: Existing low - light image enhancement methods based on diffusion models rely heavily on paired training data, which limits their wide applicability.
2. **Insufficient ability to handle unknown degradations**: Existing unsupervised methods lack effective bridging capabilities in dealing with unknown degradations, resulting in generated images often having problems such as excessive noise, color distortion and decreased visual quality.
3. **Limited generation and generalization abilities**: Although methods based on diffusion models perform well in image generation, due to the random nature of the diffusion process and the dependence on supervised constraints, these algorithms are usually based on supervised training of paired data sets or use prior knowledge of pre - trained diffusion models for network optimization, and it is difficult to achieve truly unsupervised training and wide - range practical applications.
To address these challenges, the authors propose a new zero - reference illumination - estimation diffusion model (Zero - LED) for low - light image enhancement. This model reduces the dependence on paired training data through a bidirectional - constrained unsupervised diffusion training method and improves the generation ability and the generalization ability for complex real - world scenarios.
### Main contributions
1. **Unsupervised training method with bidirectional optimization**: By combining deep neural networks and diffusion models, a diffusion model for low - light image enhancement without reference images is realized, reducing the dependence on paired training data, and enhancing the generation ability and seamless bridging between normal - light and low - light domains.
2. **Semantic and frequency - domain - guided appearance reconstruction module**: Different modalities and multiple frequency - domain spaces are used to constrain the randomness of the diffusion inference process, efficiently reconstruct images and improve the perceptual effect.
3. **Extensive experimental verification**: Experimental results on multiple public data sets show that this method outperforms other state - of - the - art unsupervised methods in both quantitative and qualitative indicators and has stronger generalization ability.
### Method overview
1. **Initial optimization network**: A deep neural network is used to pre - process the input image, generating a structural image and preliminarily optimized unknown degradation factors, providing structural constraints for the diffusion process.
2. **Diffusion - based degradation model**: The discrete wavelet transform is used to extract the low - frequency information of the low - light image, reducing the consumption of computational resources, and combining the capabilities of the generation model to simulate complex degradation processes, generating a fine illumination mask to achieve significant enhancement effects.
3. **Appearance reconstruction module**: Combining multi - modal semantic guidance and frequency - domain guidance, the reconstruction of the image content structure is guided by multiple loss functions (such as similarity loss, content loss, spectral loss, etc.), improving the quality of the generated image.
### Experimental results
The paper conducted experiments on multiple benchmark data sets, including LSRW, LOLv1, LIME and Backlit300. The experimental results show that Zero - LED performs excellently in multiple indicators such as PSNR, SSIM, NIQE and LOE, especially achieving the lowest score in the no - reference indicator NIQE, demonstrating its strong generalization ability in practical applications.
In conclusion, this paper proposes an innovative unsupervised low - light image enhancement method. Through the bidirectional - optimized diffusion model and the multi - modal - guided appearance reconstruction module, it effectively solves the limitations of existing methods and provides a new perspective for the development of low - light image enhancement techniques.