无线电工程2024，Vol.54Issue(10)：2339-2346,8.DOI:10.3969/j.issn.1003-3106.2024.10.008
基于边缘优先填充的自适应深度图像修复方法
Adaptive Depth Image Inpainting Method Based on Edge-first Filling
摘要
Abstract
To solve the problems of object edge distortion and blur caused by traditional depth image hole inpainting methods and the slow speed of repairing large holes,an adaptive depth image inpainting method based on edge-first filling is proposed.Firstly,multi-channel detection is used to extract the edges of the Red,Green,Blue and Depth(RGBD)image,and after removing the false edges of the holes and useless detail information,a salient edge of the object is generated;then,this edge is introduced into the image inpainting process,and the edge position of the hole area is filled first to effectively solve the problem of edge blurring and make the edge structure of the repaired depth image clear;finally,a gradient guidance function is introduced into the diffusion term of the Curvature Driven Diffusion(CDD)model,allowing the model to adaptively select different diffusion directions and diffusion strengths in the flat area and edge area of the hole to achieve accurate filling of large hole areas.Experimental results show that the Peak Signal to Noise Ratio(PSNR)and Mean Structural Similarity(MSSIM)of the proposed method are improved by 8～13 dB and 0.009 9～0.021 4 respectively compared with other methods on the RGBZ dataset.The proposed method can effectively repair large holes while improving iteration efficiency,and maintain clear and complete object edge contour information.关键词
深度图像/空洞修复/边缘提取/曲率驱动扩散模型/自适应扩散Key words
depth image/hole inpainting/edge extraction/CDD model/adaptive diffusion分类
信息技术与安全科学引用本文复制引用
孙梦欣,牟琦,夏蕾,李洪安,李占利..基于边缘优先填充的自适应深度图像修复方法[J].无线电工程,2024,54(10):2339-2346,8.基金项目
陕西省自然科学基础研究计划(2023-JC-YB-517) （2023-JC-YB-517）
北京航空航天大学虚拟现实技术与系统国家重点实验室开放项目(VRLAB2023B08)Natural Science Basis Research Program of Shaanxi Province(2023-JC-YB-517) （VRLAB2023B08）
Open Project of State Key Laboratory of Virtual Reality Technology and Systems of Beihang University(VRLAB2023B08) （VRLAB2023B08）
