Infrared-guided generative network for paired low-light image enhancement

Low-light image enhancement, particularly in aerial analysis, confronts significant challenges such as noise amplification, color distortion, and loss of detail. To address these issues, this paper proposes a low-light image enhancement framework based on the Brownian Bridge Diffusion Model (BBDM). This framework innovatively leverages registered Infrared (IR) imagery as auxiliary conditional guidance to generate high-quality normal-light images from their low-light counterparts. Specifically, we design a lightweight conditional encoder to efficiently extract illumination-invariant structural and textural features from the IR image. These features are subsequently fused with the image's latent space representations within the BBDM's core denoising network via across-attention mechanism. This cross-modal guidance effectively constrains the model's generative space, significantly improving detail recovery in severely degraded regions while actively suppressing the formation of artifacts. Furthermore, to enhance the sampling efficiency of the diffusion model, we introduce an adaptive step-predictor. This predictor dynamically adjusts the required number of reverse sampling steps based on the degradation level of the input image, thereby substantially reducing inference time without compromising generation quality. Comprehensive experiments were conducted on the VisDrone-vehicle dataset. By simulating varying degrees of degradation, we constructed a paired dataset comprising low-light, normal-light, and IR images. Experimental results demonstrate that our method not only effectively restores global brightness and local contrast but, more importantly, achieves significant improvements in structural fidelity and detail clarity under the precise guidance of IR information. Compared to baseline models, our approach exhibits superior artifact suppression and detail preservation, especially in severely low-light scenarios. This study confirms the substantial potential of integrating multi-modal IR information to guide generative models for low-light enhancement, paving the way for the development of efficient, high-fidelity image enhancement solutions for complex environments.