This study presents a learning-based method, DescatterNet, for real-time, non-invasive, incoherent imaging through dense and dynamic scattering media. The method is capable of imaging through turbid water and natural fog, and outperforms existing approaches in multiple aspects. The technique leverages deep learning to effectively handle natural and complex scenes, and has been tested in both indoor and outdoor environments. The results show that DescatterNet achieves outstanding performance in terms of reconstructed image quality, inference speed, and memory consumption compared to traditional image enhancement methods. The method addresses three critical challenges: acquiring "real" scattering datasets, ensuring generalization to previously unseen real-world objects, and identifying the optimal neural network architecture. The DescatterNet is trained on a dataset of real-world objects displayed on an e-ink display, and has been shown to effectively recover previously unseen real-world objects from raw images captured through the same scattering medium. The method also demonstrates cross-concentration and cross-media generalization, showing its ability to reconstruct images through different types of scattering media, including milk suspension and artificial fog. The performance of DescatterNet is evaluated under varying concentrations of scattering media, and it is found that the method's effectiveness is limited when the optical thickness of the medium exceeds a certain threshold. However, the method is capable of reconstructing images through different scattering conditions, including outdoor fog, demonstrating its robustness and flexibility in handling a wide range of real-world scenarios. The method also incorporates a preprocessing strategy that reduces the domain gap between different scattering conditions, enhancing the network's ability to generalize across various conditions. The preprocessing method includes Retinex-based processing and Contrast Limited Adaptive Histogram Equalization (CLAHE), which help to normalize the dynamic range and improve the quality of the reconstructed images. The DescatterNet is compared with several other learning-based methods, including HNN, MulScaleCNN, Unet, AttentionU-Net, and SwinIR, and is found to outperform them in terms of image quality, computational efficiency, and inference speed. The method is also shown to be effective in outdoor imaging through natural fog, demonstrating its practicality for real-world applications. The study concludes that DescatterNet is a versatile learning-based method for real-time incoherent imaging through dynamic scattering media, with the potential for practical use in a wide range of applications. The method has been demonstrated in both controlled laboratory settings and unpredictable outdoor environments, including conditions where the scattering medium and real-world objects were not part of the training data. The results suggest that DescatterNet outperforms other prevalent learning-based and traditional methods across crucial metrics, making it a practical solution for real-time applications.This study presents a learning-based method, DescatterNet, for real-time, non-invasive, incoherent imaging through dense and dynamic scattering media. The method is capable of imaging through turbid water and natural fog, and outperforms existing approaches in multiple aspects. The technique leverages deep learning to effectively handle natural and complex scenes, and has been tested in both indoor and outdoor environments. The results show that DescatterNet achieves outstanding performance in terms of reconstructed image quality, inference speed, and memory consumption compared to traditional image enhancement methods. The method addresses three critical challenges: acquiring "real" scattering datasets, ensuring generalization to previously unseen real-world objects, and identifying the optimal neural network architecture. The DescatterNet is trained on a dataset of real-world objects displayed on an e-ink display, and has been shown to effectively recover previously unseen real-world objects from raw images captured through the same scattering medium. The method also demonstrates cross-concentration and cross-media generalization, showing its ability to reconstruct images through different types of scattering media, including milk suspension and artificial fog. The performance of DescatterNet is evaluated under varying concentrations of scattering media, and it is found that the method's effectiveness is limited when the optical thickness of the medium exceeds a certain threshold. However, the method is capable of reconstructing images through different scattering conditions, including outdoor fog, demonstrating its robustness and flexibility in handling a wide range of real-world scenarios. The method also incorporates a preprocessing strategy that reduces the domain gap between different scattering conditions, enhancing the network's ability to generalize across various conditions. The preprocessing method includes Retinex-based processing and Contrast Limited Adaptive Histogram Equalization (CLAHE), which help to normalize the dynamic range and improve the quality of the reconstructed images. The DescatterNet is compared with several other learning-based methods, including HNN, MulScaleCNN, Unet, AttentionU-Net, and SwinIR, and is found to outperform them in terms of image quality, computational efficiency, and inference speed. The method is also shown to be effective in outdoor imaging through natural fog, demonstrating its practicality for real-world applications. The study concludes that DescatterNet is a versatile learning-based method for real-time incoherent imaging through dynamic scattering media, with the potential for practical use in a wide range of applications. The method has been demonstrated in both controlled laboratory settings and unpredictable outdoor environments, including conditions where the scattering medium and real-world objects were not part of the training data. The results suggest that DescatterNet outperforms other prevalent learning-based and traditional methods across crucial metrics, making it a practical solution for real-time applications.