%0 Journal Article %T 基于DenseNet多尺度融合的水下湍流鬼成像计算方法研究
Research on Underwater Turbulence Ghost Imaging Computation Method Based on DenseNet Multi-Scale Fusion %A 刘清越 %A 谢超 %J Modeling and Simulation %P 771-782 %@ 2324-870X %D 2025 %I Hans Publishing %R 10.12677/mos.2025.144328 %X 水下湍流环境对成像质量提出了严峻挑战，传统的水下成像方法在湍流影响下常常无法获得清晰图像。为了解决这一问题，本文提出了一种基于DenseNet的多尺度融合水下湍流鬼成像方法，通过深度学习优化鬼成像的图像重建精度与稳定性。与传统计算鬼成像方法(GI)及基于压缩感知的重建方法(CSGI)相比，DenseNet方法在低采样率条件下表现出明显的优势，能够在湍流干扰下恢复更为清晰和准确的图像。实验结果表明，DenseNet通过其独特的多尺度特征提取与融合能力，成功克服了湍流对光传播造成的失真与噪声，提高了图像的细节保真度和结构恢复效果。该方法不仅展示了在低采样率下的强大恢复能力，还为复杂水下环境中的成像技术提供了新的解决方案，推动了基于深度学习的水下湍流成像技术的进一步发展。
Underwater turbulence presents significant challenges to imaging quality, and traditional underwater imaging methods often fail to produce clear images under turbulent conditions. To address this issue, this paper proposes a multi-scale fusion underwater turbulence ghost imaging method based on DenseNet, which enhances the accuracy and stability of ghost imaging reconstruction through deep learning. Compared with traditional ghost imaging methods (GI) and compressed sensing-based reconstruction methods (CSGI), the DenseNet method demonstrates significant advantages under low sampling rates, enabling clearer and more accurate image restoration even in the presence of turbulence. Experimental results show that DenseNet successfully overcomes distortion and noise caused by underwater turbulence through its unique multi-scale feature extraction and fusion capabilities, improving the fidelity and structural recovery of the image details. This method not only exhibits strong recovery capabilities under low sampling rates but also provides a new solution for imaging technologies in complex underwater environments, advancing the development of deep learning-based underwater turbulence imaging technologies. %K 水下湍流， %K 鬼成像， %K DenseNet， %K 多尺度融合， %K 图像重建， %K 深度学习， %K 湍流建模
Underwater Turbulence %K Ghost Imaging %K DenseNet %K Multi-Scale Fusion %K Image Reconstruction %K Deep Learning %K Turbulence Modeling %U http://www.hanspub.org/journal/PaperInformation.aspx?PaperID=112523