基于多特征重构的三维目标反演算法

Three-dimensional target inversion algorithm based on multi-feature reconstruction

为了解决基于深度学习的三维反演方法中存在的内存占用大、训练耗时久的问题,提出基于多特征重构的三维目标反演算法.通过特征分解提取目标的水平区域、中心深度、垂直厚度和剩余密度4类特征,实现对三维模型的压缩,降低内存占用.设计多特征重构反演网络(MRNet),通过不同的Decoder实现对目标4类特征的预测,利用4类特征重构三维模型,实现对三维目标的反演.在网络输入端引入梯度联合实现对目标边界信息的增强.在跨层连接处引入CA注意力机制,实现对Decoder预测功能的分化,优化反演效果.模拟实验结果显示,MRNet的局部相对准确度相对于3D U-Net提升了30%以上,达到88.91%,每轮训练时间仅为3D U-Net的1/13.将MRNet应用于Vinton盐丘地区,较准确地得到了盖岩的分布情况,验证了MRNet具备一定的泛化性.

A 3D target inversion algorithm based on multi-feature reconstruction was proposed in order to solve the problems of large memory occupation and time-consuming training in deep learning-based three-dimensional inversion methods.Four types of features,horizontal area,center depth,vertical thickness and residual density of the target were extracted by feature decomposition to realize the compression of the three-dimensional model and reduce the memory occupation.The multi-feature reconstruction of inversion network(MRNet)was designed to realize the prediction of the four types of target features by different Decoder,and the four types of features were used to reconstruct the three-dimensional model to realize the inversion of the 3D target.The gradient union was introduced at the input of the network to realize the enhancement of target boundary information.The CA attention mechanism was introduced at the cross-layer connection to realize the differentiation of Decoder’s prediction function and optimize the inversion effect.The simulation results showed that the local relative accuracy of MRNet was improved by more than 30%compared with 3D U-Net,reaching 88.91%,and the training time per round was only 1/13 of 3D U-Net.MRNet was applied to Vinton Salt Mound,and the distribution of caprocks was obtained more accurately,which verified that MRNet had certain generalizability.