Yue Zelin, Ruan Ping, Chen Songmao, Fang Mengyan, Su Yongyi, Liu Hui, Chen Minglai, Wang Xing, Yang Xulei, Veeravalli Bharadwaj, Xu Xun
Opt Express. 2025 Jul 28;33(15):31042-31061. doi: 10.1364/OE.562113.
Accurate and robust 3D reconstruction for underwater streak tube imaging LiDAR (STIL) is hindered by weak echo signals and poor imaging quality due to high scattering, absorption, and turbidity. Traditional signal processing-based enhancement methods often neglect the semantic and distributional information of reconstructed objects. To address this limitation, we propose an RGBD diffusion model for denoising reconstructed images via generative refinement. Given the scarcity of underwater training data, we pre-train the diffusion model on an external dataset. To further enforce consistency between the generated content and the input, we incorporate a state-of-the-art cross-correlation algorithm (CCA) to guide the low-frequency components during the diffusion process. Experimental results demonstrate that our approach achieves higher accuracy and lower errors compared to existing methods, with a depth resolution surpassing 0.55 mm under a 0.14 m water attenuation length, significantly enhancing underwater STIL imaging performance.
水下条纹管成像激光雷达（STIL）的精确且稳健的三维重建受到弱回波信号以及由于高散射、吸收和浑浊导致的成像质量差的阻碍。传统的基于信号处理的增强方法常常忽略重建物体的语义和分布信息。为了解决这一局限性，我们提出了一种RGBD扩散模型，通过生成式细化对重建图像进行去噪。鉴于水下训练数据的稀缺性，我们在外部数据集上对扩散模型进行预训练。为了进一步加强生成内容与输入之间的一致性，我们纳入了一种最先进的互相关算法（CCA），以在扩散过程中引导低频分量。实验结果表明，与现有方法相比，我们的方法具有更高的精度和更低的误差，在0.14米的水衰减长度下深度分辨率超过0.55毫米，显著提高了水下STIL成像性能。
