Unsupervised lensless image reconstruction driven by structure consistency modeling in low-light scenarios

Lensless imaging offers unique advantages in device miniaturization, making it attractive for endoscopy, wearable devices, and microrobotics. However, its reliance on coded masks inherently reduces photon efficiency, and under lowlight conditions, reconstructions suffer from amplified noise, blurred boundaries, and distorted structures. To address these challenges, we introduce an unsupervised reconstruction framework driven by structure consistency modeling. At the local structure level, we design a dynamic pixel selection strategy based on partial convolution, enabling the network to selectively focus on high-SNR regions and suppress noise-dominated areas. At the global structure level, we impose a geometric equivariance constraint, requiring reconstructions to remain consistent under transformations such as rotation and mirroring. This dual-level modeling jointly enhances fine-grained detail recovery and overall structural stability. Experiments on public datasets demonstrate that our method achieves superior noise suppression and structure preservation compared with existing approaches, highlighting the effectiveness of structure consistency modeling as a principle for unsupervised low-light lensless imaging.