高斯边缘增强的自监督单目深度估计

打开文本图片集

引用格式：，，刘鹏.高斯边缘增强的自监督单目深度估计[J].现代电子技术，2026，49（7）：63-68.

关键词：自监督学习；单目深度估计；高斯金字塔；边缘增强；ASPP；语义特征中图分类号：TN911.73-34；TP391.4 文献标识码：A 文章编号：1004-373X（2026）07-0063-06

Self-supervised monocular depth estimation based on Gaussian edge enhancement

HUANG Yue’， ZHANG Peng'，LIU Peng² （1.Schoolof Instrument and Electronics，North UniversityofChina，Taiyuan O3oo51，China; 2.School of ElectricalandControl Engineering，North Universityof China，Taiyuan O3oo51，China）

Abstract：Self-supervised monoculardepth estimation methods basedonencoder-decoderarchitecturesoften suffer from blured edges indepth maps duetoupsamplingoperations.Existing solutions predominantly introduceedgeconstraints during thestagofdecodingorwithinthelossfunction，whichfaceslimitationsofposterioroptimizationafterhighfrequencyinforation atenuation.Inviewof this，theauthor proposes asource-enhanced Gausianedgeenhancement （GEE）mechanism.Thecore inovationliesin：explicitlyconstructingadiffrenceofGaussian （DoG）pyramidduringthepreprocesingstagefirsttodecouple multi-scaleedgepriors fromtheinput image;subsequently，designinganadaptiveedgeinjection （AEI）module toachieve dynamicfusionofgeometricandsemanticfeaturesatthefrontendoftheencoder;finalycombininganedge-guidedASPP ++ module toenhancecontextualawarenessExperimentsontheKITTIdatasetshowthattheRMSE，AbsRel，andSqRelof the proposed method reduce by 14.83% ， 8.92% ，and 28.O8%，respectively，in comparison with those of the mainstream algorithms.In addition，theproposed method significantlyoutperforms the latest SOTAmethodssuchasBTSandDIFFNet.Thevisualization results have verified its excellent depth discontinuity modeling ability in complex contour and weak texture areas.

Keywords：self-supervised learning；monoculardepthestimation；Gausianpyramid;edge enhancement;ASPP;semantic feature

0 引言

单目深度估计旨在从单张图像恢复场景三维结构，是自动驾驶、机器人导航等应用领域的关键技术。（剩余9987字）