自动化学报2025，Vol.51Issue(11)：2498-2519,22.DOI:10.16383/j.aas.c250024
基于深度语义扩散的深度图修复:缺陷数据集与模型
Depth Map Repair Based on Depth Semantic Diffusion:Defect Dataset and Model
摘要
Abstract
Depth repair aims to address the issues of missing data,noise and occlusion in depth maps during 3D re-construction.However,due to the diversity and heterogeneity of depth map sources,existing depth repair methods struggle to effectively repair complex scene structures and unknown types of depth defects.To address these issues,unlike current approaches that focus solely on enhancing algorithmic robustness,we propose a novel solution from the inverse perspective of depth defect dataset construction.We construct a real defect sampling simulation(RDSS)dataset.Based on this,we propose a DR-Net depth map repair model utilizing depth semantic diffusion.The RDSS dataset constructs formalized simulations of diverse depth defects in complex scenes by capturing and modeling real defects,combined with homogeneous deformation augmentation and heterogeneous cross-combination.This signific-antly enhances depth defect diversity and scene coverage.The designed DR-Net model builds upon a U-Net struc-ture.It employs a reverse transmission module to preserve high-resolution details while propagating depth semantic information within the image through a depth semantic diffusion module,thereby effectively improving repair per-formance.Experimental results demonstrate that models trained on the RDSS benchmark dataset achieve effective depth map repair across other datasets.Furthermore,compared to state-of-the-art model design repair method(SD-Filter)and data-driven repair method(G2),the DR-Net model reduces the root mean squared error metric by 24.85％and 29.54％on average across the RDSS,NYU Depth V2 and KITTI datasets,respectively.These results validate the effectiveness and advancement of the proposed DR-Net model.关键词
深度图修复/柏林噪声/真实深度缺陷采集/深度语义扩散/深度缺陷数据集Key words
Depth map repair/Perlin noise/real depth defect collection/depth semantic diffusion/depth defect dataset引用本文复制引用
闫涛,李彤,张江峰,钱宇华,陈路,吴鹏..基于深度语义扩散的深度图修复:缺陷数据集与模型[J].自动化学报,2025,51(11):2498-2519,22.基金项目
国家自然科学基金(T2495250,T2495251,62136005,62472268,62373233),中央引导地方科技发展资金项目(YDZJSX2023C001,YDZJSX2023B001)资助Supported by National Natural Science Foundation of China(T2495250,T2495251,62136005,62472268,62373233)and Funds for Central-government-guided Local Science and Techn-ology Development(YDZJSX2023C001,YDZJSX2023B001) （T2495250,T2495251,62136005,62472268,62373233）
