Multi-view Stereo Vision Reconstruction Network with Fusion Attention Mechanism and Multi-layer Dynamic Deformable Convolution
The existing multi-view stereo vision technology is not enough to extract the feature information of weak texture region and non-Lambert surface,and its reconstruction effect is not ideal.An AMDC-PatchmatchNet method with fusion attention mechanism and multi-layer dynamic deformable convolution is proposed for the problems above.In this method,a feature extraction network integrating the coordinate attention is constructed,which can capture the edge shape and texture features of reconstructed objects more accurately.At the same time,an adaptive receptive field module based on dynamic deformable convolution is integrated in the feature extraction network,and the size and shape of receptive field can be adjusted adaptively according to different scales of features to obtain both global and detailed feature representation.The generalization ability of the AMDC-PatchmatchNet method is verified on the aerial image data sets.The test results on DTU data sets show that the overall index of point cloud reconstruction of the proposed method is improved by 2.8％compared with those of mainstream MVS methods.
万方数据
维普
