Robust speckle reconstruction based on cascade transfer learning and speckle correlation imaging

Abstract

Achieving high-quality reconstruction with unknown scattering media and complex scattering conditions, such as low signal-to-noise ratio (SNR) or non-darkroom environments with strong ambient light, remains a significant challenge. Traditional imaging methods have good generalization ability but the fidelity of results needs to be improved, while deep learning methods have good imaging results but limited generalization ability. In order to enhance the generalization ability of the model, improve the reconstruction quality, and achieve robust reconstruction in high-intensity ambient light noise environments, we propose a method based on cascade transfer learning and speckle correlation imaging. Specifically, an innovative and flexible cascade transfer learning architecture is proposed for accurate and robust speckle reconstruction, while the speckle correlation imaging chain is used to generate robust pre-training and fine-tuning datasets, maximizing the advantages of the pre-training and boosting the overall efficacy of transfer learning. Additionally, a degradation-aware Transformer network is designed to achieve better convergence in both pre-training and fine-tuning tasks. Experimental results show that our method outperforms traditional methods and various deep learning-based approaches in both reconstruction fidelity and generalization. Moreover, it can reliably reconstruct targets utilizing low-quality speckles in unfavorable environments, and successfully tackle the challenge of reconstructing highly complex face images through biological tissue, offering new inspiration for scattering imaging.