Research On Salt Body Semantic Segmentation Based On Deep Learning

Seismic image analysis plays a vital role in industrial applications and has received widespread attention.One of the main challenges of seismic image analysis is detecting the salt structure underground,which is essential for identifying oil and gas reservoirs and planning drilling paths.At present,traditional seismic image analysis still requires professionals to analyze salt bodies.Convolutional neural networks have been successfully applied in many fields,and many attempts have been made in the field of seismic imaging.The data characteristics of salt bodies are different from natural images.For example,the shape of the salt body is uncertain and there is no prior knowledge of the shape;the texture of the geological data is more prominent,and so on.Therefore,the general semantic segmentation network performs poorly on salt data.Aiming at these characteristics of salt data,we designed a deep supervised semantic segmentation model,and optimized the segmentation results through edge prediction branch.The main research contents of this paper are as follows:First of all,in response to the problem of indefinite salt data,this paper designs a multi-task,deep-supervised,salt body semantic segmentation method.By constructing a classification and multi-path collaborative optimization model,it achieves effective segmentation of salt data.This improves the accuracy of the segmentation results.Secondly,in order to solve the problem of small target segmentation,an edge prediction branch is designed to predict the boundary of the salt body.It supervises the loss of the boundary to guide the feature learning,so that the network can more clearly distinguish the semantic boundary between the salt body and the non-salt body.The features on the side effectively improve the accuracy of the salt body boundary segmentation,thereby improving the segmentation performance of small targets,and ultimately improving the overall performance of salt body segmentation.Finally,in view of the limited resolution and insufficient data volume of the salt body data,this paper uses all the feature spectra of different resolutions in the salt body segmentation model decoder,and supervises it through the semantic segmentation loss function,which increases the utilization of the decoder features.In addition,the features of edge prediction branche and the features of segmentation branches are merged to improve the expression ability of the features,which effectively improves the performance of salt segmentation.Experiments show that the algorithm in this paper effectively improves the overall segmentation performance through the multi-task deep supervision method.The introduction of edge prediction branch effectively improves the segmentation accuracy of small targets.At the same time,the multi-scale supervised salt body segmentation method and the fusion of boundary features effectively improve the salt body segmentation performance.