Unveiling Advanced Frequency Disentanglement Paradigm for Low-Light Image Enhancement

Kun Zhou,Xinyu Lin,Wenbo Li,Xiaogang Xu,Yuanhao Cai,Zhonghang Liu,Xiaoguang Han,Jiangbo Lu
2024-09-03
Abstract:Previous low-light image enhancement (LLIE) approaches, while employing frequency decomposition techniques to address the intertwined challenges of low frequency (e.g., illumination recovery) and high frequency (e.g., noise reduction), primarily focused on the development of dedicated and complex networks to achieve improved performance. In contrast, we reveal that an advanced disentanglement paradigm is sufficient to consistently enhance state-of-the-art methods with minimal computational overhead. Leveraging the image Laplace decomposition scheme, we propose a novel low-frequency consistency method, facilitating improved frequency disentanglement optimization. Our method, seamlessly integrating with various models such as CNNs, Transformers, and flow-based and diffusion models, demonstrates remarkable adaptability. Noteworthy improvements are showcased across five popular benchmarks, with up to 7.68dB gains on PSNR achieved for six state-of-the-art models. Impressively, our approach maintains efficiency with only 88K extra parameters, setting a new standard in the challenging realm of low-light image enhancement.
Computer Vision and Pattern Recognition
What problem does this paper attempt to address?
### What problem does this paper attempt to solve? This paper aims to solve the problem of frequency - coupling optimization in low - light image enhancement (LLIE). Specifically, existing LLIE methods usually handle low - frequency adjustment (such as illumination restoration) and high - frequency restoration (such as noise reduction) in a unified framework, which leads to the problems of complexity and sub - optimal results. #### Main problems: 1. **Challenges of frequency - coupling optimization**: - Existing LLIE models usually adopt a single, complex network structure when dealing with low - and high - frequency problems, which makes the optimization process very difficult. - For example, low - frequency adjustment may inadvertently amplify noise, while high - frequency restoration may affect the restoration of illumination intensity. 2. **Lack of decoupling in the optimization process**: - Existing frequency - decomposition learning schemes ignore the optimization interaction between different frequency components and need to handle multiple degradation problems simultaneously. - This coupled optimization may lead to inaccurate illumination correction or residual noise artifacts. 3. **How to design a general frequency - decoupling paradigm**: - How to design a frequency - decoupling paradigm that can be seamlessly integrated into existing LLIE methods to enhance their frequency - restoration ability with minimal additional model complexity (such as additional parameters and computational cost). ### Solutions: To solve the above problems, the authors propose a new frequency - decoupling optimization paradigm. The main contributions include: 1. **Introducing an effective frequency - decoupling learning framework**: - Decompose the complex coupled - optimization problem into two relatively simple tasks: rough low - frequency adjustment and frequency - decoupled fine - restoration. - Use Laplacian pyramid to decompose the image and propose a new low - frequency consistency constraint to ensure that low - frequency information remains consistent between the two stages. 2. **Efficient rough - adjustment module (ACCA)**: - Propose an Adaptive Convolutional Composition Aggregation module (ACCA) for rough low - frequency adjustment. - ACCA effectively reduces the overall complexity by designing a convolutional - style spatial - channel attention mechanism and achieves a significant PSNR improvement on the LOL - v2 dataset (1 dB higher than Retinexformer). 3. **Laplacian - decoupled restoration model (LDRM)**: - LDRM is mainly used for high - frequency enhancement. By introducing a new low - frequency consistency term, it greatly simplifies the optimization process. - It shows significant improvement on multiple benchmark datasets. For example, it improves the PSNR of Restormer by 7.68 dB on the SDSD - out benchmark. 4. **Extensive improvement effects**: - This method not only has a significant improvement in quantitative indicators but also has an obvious improvement in visual quality, effectively solving the problems of illumination degradation and artifacts. Through these innovations, this paper provides an efficient and general - purpose solution that can significantly improve the performance of existing LLIE models with minimal additional computational cost.