高斯边缘增强的自监督单目深度估计OA

Self-supervised monocular depth estimation methods based on encoder-decoder architectures often suffer from blurred edges in depth maps due to upsampling operations.Existing solutions predominantly introduce edge constraints during the stage of decoding or within the loss function,which faces limitations of posterior optimization after high-frequency information attenuation.In view of this,the author proposes a source-enhanced Gaussian edge enhancement(GEE)mechanism.The core innovation lies in:explicitly constructing a difference of Gaussian(DoG)pyramid during the preprocessing stage first to decouple multi-scale edge priors from the input image;subsequently,designing an adaptive edge injection(AEI)module to achieve dynamic fusion of geometric and semantic features at the front end of the encoder;finally,combining an edge-guided ASPP++module to enhance contextual awareness.Experiments on the KITTI dataset show that the RMSE,Abs Rel,and Sq Rel of the proposed method reduce by 14.83%,8.92%,and 28.08%,respectively,in comparison with those of the mainstream algorithms.In addition,the proposed method significantly outperforms the latest SOTA methods such as BTS and DIFFNet.The visualization results have verified its excellent depth discontinuity modeling ability in complex contour and weak texture areas.