Research On Feature Instance Extraction Technique For Imaging Iogging Images

In research in the field of petroleum exploration,the effective extraction of geological features is crucial to the task of interpretation and analysis of imaging log images.However,because most of the geological features depicted in imaging logging images are characterized by weak boundaries(transition regions with pixel blurring at the junction of geological feature regions and background regions),it is difficult to achieve more accurate description results by using only traditional image segmentation techniques.In recent years,deep learning methods have been widely used in segmentation tasks of imaging logging images,and have achieved excellent results in many scenarios,surpassing the performance of traditional methods.However,such methods still have shortcomings in the face of weak boundary scenarios.As a result,it is often difficult for existing general-purpose methods to finely capture the edge information of geological features,which results in deviations of segmentation results from the true labels.To solve this problem,this study designs a semantic segmentation network named Bilateral FS-UNet for complex scenarios of imaging logging.the main contributions are as follows:(1)A filtering technique(BFM)is designed for complex scenarios of imaging logging because of the prevalence of weak boundaries in geological features in imaging logging images,which makes it difficult to acquire geological features accurately.The technique is able to enhance the boundaries of geological feature areas to maximize the elimination of weak boundaries and the effects they bring;(2)An adaptive feature stimulation module(AFSM)is designed to cope with the processing needs of complex geological features(such as spatial intersection with other geological features,etc.)in imaging logging images.This module aims to solve the mis-segmentation cases caused by spatial intersection and other reasons during segmentation,and thus improve the segmentation accuracy.It is verified that the average pixel accuracy of Bilateral FS-UNet is able to reach97.21%,which is 2.09% better than the general segmentation model U-Net.This indicates that the network performs well for coping with weak boundary problems in imaging logging scenarios and is expected to enhance the automatic segmentation and interpretation of geological features in practical applications.