The objective function of full waveform inversion is a strong nonlinear function,the inversion process is not unique,and it is easy to fall into local minimum.Firstly,in the process of wavefield reconstruction,the wav...The objective function of full waveform inversion is a strong nonlinear function,the inversion process is not unique,and it is easy to fall into local minimum.Firstly,in the process of wavefield reconstruction,the wave equation is introduced into the construction of objective function as a penalty term to broaden the search space of solution and reduce the risk of falling into local minimum.In addition,there is no need to calculate the adjoint wavefield in the inversion process,which can significantly improve the calculation efficiency;Secondly,considering that the total variation constraint can effectively reconstruct the discontinuous interface in the velocity model,this paper introduces the weak total variation constraint to avoid the excessive smooth estimation of the model under the strong total variation constraint.The disadvantage of this strategy is that it is highly dependent on the initial model.In view of this,this paper takes the long wavelength initial model obtained by first arrival traveltime tomography as a prior model constraint,and proposes a weak total variation constrained wavefield reconstruction inversion method based on first arrival traveltime tomography.Numerical experimental results show that the new method reduces the dependence on the initial model,the interface description is more accurate,the error is reduced,and the iterative convergence efficiency is significantly improved.展开更多
为了提升弱纹理区域无监督多视图深度估计性能,文中提出一种基于邻域自适应无监督多视图深度估计算法。算法采用双分支结构,深度估计分支首先采用邻域自适应深度分布方法改善弱纹理区域深度分布;其次采用深度变化概率引导的深度假设范...为了提升弱纹理区域无监督多视图深度估计性能,文中提出一种基于邻域自适应无监督多视图深度估计算法。算法采用双分支结构,深度估计分支首先采用邻域自适应深度分布方法改善弱纹理区域深度分布;其次采用深度变化概率引导的深度假设范围细化后续阶段深度估计。为了提高对场景边缘的识别,采用基于标准差的深度平滑约束。神经渲染分支用于提高深度估计能力,为了增强与深度估计分支间的几何一致性,采用融合图像颜色与深度信息的采样方法。由实验结果可知,该算法在DTU数据集测试完整度误差和整体精度误差优于其他无监督算法,且完整度误差比DS⁃MVSNet减小16.71%。可视化结果表明,针对弱纹理区域深度估计性能提升明显。在Tanks and Temples数据集上进行泛化性验证,整体性能(Mean)为56.22,证明了所提算法的有效性。展开更多
基金supported by National Key R&D Program of China under contract number 2019YFC0605503CThe Major projects of CNPC under contract number(ZD2019-183-003)+2 种基金the Major projects during the 14th Five-year Plan period under contract number 2021QNLM020001the National Outstanding Youth Science Foundation under contract number 41922028the Funds for Creative Research Groups of China under contract number 41821002.
文摘The objective function of full waveform inversion is a strong nonlinear function,the inversion process is not unique,and it is easy to fall into local minimum.Firstly,in the process of wavefield reconstruction,the wave equation is introduced into the construction of objective function as a penalty term to broaden the search space of solution and reduce the risk of falling into local minimum.In addition,there is no need to calculate the adjoint wavefield in the inversion process,which can significantly improve the calculation efficiency;Secondly,considering that the total variation constraint can effectively reconstruct the discontinuous interface in the velocity model,this paper introduces the weak total variation constraint to avoid the excessive smooth estimation of the model under the strong total variation constraint.The disadvantage of this strategy is that it is highly dependent on the initial model.In view of this,this paper takes the long wavelength initial model obtained by first arrival traveltime tomography as a prior model constraint,and proposes a weak total variation constrained wavefield reconstruction inversion method based on first arrival traveltime tomography.Numerical experimental results show that the new method reduces the dependence on the initial model,the interface description is more accurate,the error is reduced,and the iterative convergence efficiency is significantly improved.
文摘为了提升弱纹理区域无监督多视图深度估计性能,文中提出一种基于邻域自适应无监督多视图深度估计算法。算法采用双分支结构,深度估计分支首先采用邻域自适应深度分布方法改善弱纹理区域深度分布;其次采用深度变化概率引导的深度假设范围细化后续阶段深度估计。为了提高对场景边缘的识别,采用基于标准差的深度平滑约束。神经渲染分支用于提高深度估计能力,为了增强与深度估计分支间的几何一致性,采用融合图像颜色与深度信息的采样方法。由实验结果可知,该算法在DTU数据集测试完整度误差和整体精度误差优于其他无监督算法,且完整度误差比DS⁃MVSNet减小16.71%。可视化结果表明,针对弱纹理区域深度估计性能提升明显。在Tanks and Temples数据集上进行泛化性验证,整体性能(Mean)为56.22,证明了所提算法的有效性。
