Retinex based Dual-End Specialized Diffusion Model for Low-Light Image Enhancement

Among the various solutions, image enhancement methods based on physical models have strong interpretability but are somewhat lacking in detail and texture. Image enhancement methods that employ generative models can significantly recover image details. However, they often lack reliable fidelity constraints, tending to synthesize details that are visually plausible but rather than physically accurate. This tendency can lead to alterations in the semantic information of the original scene. To integrate the advantages of both physical model and generative model, a diffusion-based low-light image enhancement algorithm is proposed. Initially, decompose the images into reflection components and illumination components, allowing for more targeted processing of different domains within the images. Subsequently, a stable diffusion model is applied to the reflection component for denoising and detail recovery. Within the stable diffusion model, a degradation removal module and a parallel control module are introduced to direct the generation of the model, ensuring the fidelity of the generated images, while hidden space encoding and cross-attention mechanisms are employed to mitigate color shift issues. Following this, the illumination component employs light-guided conditional diffusion to enhance the brightness and contrast of the image while controlling the light intensity to conform to human visual. Finally, the processed reflection components and illumination components are fused to obtain the final enhanced result. The proposed method effectively addresses various degradation phenomena such as noise and artifacts that occur during low-light image processing, resulting in the production of refined images with consistent content and robustness.