A Low-Light Image Enhancement Algorithm Based on Improved Pyramid Diffusion Model

Abstract Images captured in low-light conditions often suffer from poor visibility, high noise levels, and significant detail loss. These issues can severely hinder subsequent visual tasks like object detection and facial recognition. Therefore, low-light image enhancement is a crucial yet challenging problem in computer vision, aiming to recover high-quality images.Pyramid-based Diffusion Model, a type of generative model capable of modeling and generating high-dimensional data distributions, have recently been explored for low-light image enhancement. However, a common limitation of diffusion models is the potential loss of detail in small areas due to the forward noising and reverse denoising process. To address this, the proposed algorithm integrates a novel attention mechanism that progressively utilizes depthwise separable convolution units with expanding convolutional kernels. This approach effectively enhances the network's ability to handle larger receptive fields, leading to better detail preservation. Furthermore, the algorithm incorporates a dual interpolation sampling method for improved detail recovery. Additionally, a new combination optimization model for loss functions is adopted. Experimental results demonstrate that the proposed method outperforms the baseline network (which lacks color correction methods) on both subjective and objective evaluation metrics across LOL-v1, LOL-v2, and other unpaired datasets. This performance surpasses current mainstream low-light enhancement algorithms.