Classical image formation assumes minimal light distortion during propagation. However, scattering media (biological tissues, haze, fog, turbid water) severely degrade images due to scattering noise. Traditional methods focus on isolating early-arriving light or improving signal-to-noise ratio, using techniques like gating methods leveraging the Kerr effect or coherence/polarization characteristics. These methods are limited by the exponential decay of ballistic light with optical thickness. Computational techniques, particularly deep learning (DL), offer an alternative by utilizing both early and late-arriving scattered light. However, existing DL approaches often rely on artificial, static scattering media and non-representative training datasets (e.g., images from MNIST, CIFAR, ImageNet displayed on SLMs/DMDs), limiting their real-world applicability. This paper introduces DescatterNet, a learning-based method designed to address these limitations by using real-world datasets and a robust neural network architecture to enable imaging through dynamic and inhomogeneous scattering media.

Existing research on imaging through scattering media has explored various approaches. Traditional methods focused on isolating early-arriving ballistic photons from multiply-scattered light, employing techniques such as time-gating using the Kerr effect, or leveraging coherence and polarization properties. These methods, while innovative, suffer from limitations in depth penetration due to the exponential decay of ballistic photons with increasing optical thickness. Computational methods, particularly deep learning, have emerged as a promising alternative, capable of leveraging both ballistic and scattered light components. However, these deep learning-based approaches have typically relied on artificial, static scattering media and synthetic training datasets, hindering their real-world application and generalization capabilities. This work addresses these shortcomings by proposing a novel method that employs a more realistic and practical approach.

DescatterNet addresses three key challenges: acquiring realistic scattering datasets, ensuring generalization to unseen real-world objects, and optimizing the neural network architecture. First, a custom experimental setup was used to collect thousands of "real" scattered-clear image pairs under various dynamic scattering conditions (turbid water, fog). Second, a pre-processing method bridges the domain gap between training data (images displayed on an e-ink display) and real-world objects, improving generalization. Third, an optimized neural network architecture was developed and compared to existing alternatives (HNN, MulScaleCNN, U-Net, AttentionUNet, SwinIR). The network's architecture was selected through experimentation and a comparison of several different models, with the focus on parameters, computational complexity, inference speed and image quality metrics. The selection criteria were focused on obtaining the best results in terms of image quality and speed. The training process involved using a dataset collected using a customized setup designed to collect real scattered and clear images in various scattering conditions. A pre-processing step was used to improve the generalization capability of the model, bridging the gap between the training dataset and the real-world scenarios. Various neural network architectures were considered and compared before selecting the final architecture for DescatterNet.

DescatterNet significantly outperforms traditional methods (dark channel prior, Retinex) and other deep learning-based approaches (MulScaleCNN) in reconstructing images from highly scattered data. The quantitative results (PSNR, Correlation Coefficient) demonstrate superior image quality. Experiments showed that DescatterNet successfully recovers previously unseen real-world objects through dynamic scattering media (turbid fat emulsion, fog). Analysis of the upper limit of descattering performance indicated that image quality deteriorates with increasing optical thickness, with a critical point observed around a specific fat emulsion volume. The cross-concentration generalization capability of DescatterNet was evaluated, showing its ability to handle varying scattering densities without retraining. Comparative analysis across various neural network architectures revealed that DescatterNet offers a favorable balance between model size, computational complexity, inference speed, and image reconstruction quality, outperforming the other examined networks in terms of accuracy and efficiency.

DescatterNet's success in real-time imaging through dynamic scattering media demonstrates the power of combining sophisticated deep learning techniques with carefully designed data acquisition methods. The superior performance compared to traditional and other deep learning methods highlights the effectiveness of the proposed approach in handling the complexities of real-world scattering scenarios. The identified upper limit on the effective optical thickness suggests future research directions focusing on improved data acquisition techniques and network architectures to further enhance performance. The ability of DescatterNet to generalize across varying scattering densities underscores its robustness and potential for broad applicability in diverse imaging applications.

This study presents DescatterNet, a novel learning-based method for real-time imaging through dynamic scattering media. DescatterNet surpasses existing methods in image quality, speed, and generalizability. Future work could focus on extending the method's capabilities to even denser media and exploring its applications in various fields like biomedical imaging and remote sensing.

The current implementation of DescatterNet has an upper limit on the optical thickness of the scattering media it can effectively penetrate. Further research is needed to determine if this is an intrinsic limitation or if improvements in data acquisition or network architecture could overcome this constraint. The training data, while more representative than previously used datasets, may still not fully capture the diversity of real-world scattering scenarios. The generalization performance across different types of scattering media was not explicitly assessed.