Imaging Through Scattering Media With Speckle-Correlation-Based Learning

Imaging through scattering media has significant applications in fields such as optical communication and astronomical observation. In recent years, deep learning (DL) has achieved breakthrough progress in this field due to its powerful capabilities in feature extraction and nonlinear fitting, offering new solutions to the challenges of imaging through scattering media. However, traditional DL methods typically require large-scale datasets and exhibit limited generalization capabilities, often performing well only under specific distributions of scattering media. When the properties of the scattering media change, model performance tends to degrade significantly, which severely restricts practical applicability. In this paper, we propose a method that combines speckle-correlation theory with DL. By introducing this theory, the model can learn additional physical prior to improve generalization ability. Experimental results show that the proposed method can accurately reconstruct ground truth objects under different scattering media, and remains effective for arbitrary complex grayscale input. These results validate both the accuracy and generalization ability of the method. This study demonstrates the effectiveness of combining speckle-correlation theory with DL to overcome the limited generalization of data-driven methods, providing a more robust and scalable solution to scattering imaging, particularly in dynamic and complex scattering media.