Accurate Unsupervised Stereo Matching Algorithm for Ill-Posed Regions Based on CTH
The accuracy of unsupervised stereo matching algorithms directly affects the quality of depth estimation.Current unsupervised stereo matching algorithms generally lack the ability to extract global features,resulting in low disparity estimation accuracy in ill-posed regions such as edges and occlusions.To address this issue,an accurate unsupervised stereo matching algorithm based on CNN and Transformer for global and local feature extraction in ill-posed regions is proposed.First,a sandglass network combining multi-scale and skipping connections is intro-duced,and a CNN-Transformer Hybrid(CTH)algorithm is designed for stereo matching to enable parallel feature extraction.Second,a Feature Coupling Unit(FCU)is used to fuse local features with global representations,en-hancing the model's ability to integrate global information.Finally,the original disparity supervision loss function is incorporated into the baseline model's loss function,guiding the network to learn more comprehensive features through consistency loss.Experimental results demonstrate that compared to other classical algorithms,the proposed method improves accuracy and robustness in ill-posed regions.
unsupervised stereo matchingCNNTransformerCTHglobal and local feature extraction
国家科技期刊平台
NSTL
万方数据
