Research On Rock Fracture Segmentation In Outcrop Area Based On Improved Deeplabv3+

Rock fractures in outcrop area provide important data support for oil and gas geological exploration,which is the basic research content of oil and gas exploration and development.In the process of geological exploration in the field,the survey of rock strata and fracture distribution in the outcrop area is one of the most important links.Manual sketching is a common mapping method,but this method has the characteristics of low efficiency and strong subjectivity.How to accurately segment rock cracks from different rock images in the field outcrop area has always been a difficult point.The shape and size of the cracks are different,and the distribution position in the rock is different,and the rock image has some small stones and other debris,illumination shadow and other problems.The traditional image segmentation algorithm cannot achieve good segmentation results.This paper proposes a semantic segmentation model based on Deeplabv3+,which realizes the accurate segmentation and rock classification of fractures in the background of dolomite,limestone,mudstone and sandstone in the field outcrop area.In this paper,the rock images of the field outcrop area are collected,and the image data are screened,enhanced contrast,labeling and other image processing work,and the data set that can be used for model training in this paper is obtained.Aiming at the problem of large number of parameters and poor segmentation effect of the original Deeplabv3+ network,an optimized network structure was designed.Firstly,the improved model uses a more lightweight Mobilenetv2 to replace the Xception network with a larger number of original parameters as the backbone network of the model,which reduces the number of parameters of the model.Secondly,a cross-layer connection operation is used between different hole convolutions of the ASPP structure,so that the back layer can use the feature information of different scales of the front layer to enhance the feature extraction ability of the network,improve the connectivity and integrity of the rock fracture segmentation,reduce the misclassification and improve the edge effect of the fracture segmentation.Thirdly,the information of the channel attention module is introduced into the spatial attention module,and an improved L-CBAM attention module is proposed.After the feature extraction network,it is added in parallel with ASPP to the model,which enhances the network ’s ability to obtain global context feature information and improves the segmentation accuracy of the model for crack edge details.Finally,in view of the unbalanced number of samples in different categories in the data of this paper,the Focal loss function is used to improve the problem of unbalanced segmentation of different categories and improve the segmentation accuracy of crack types.Based on the above aspects,a semantic segmentation model with encoder decoder structure was optimized for Deeplabv3+network,achieving accurate segmentation of different rocks and fractures such as dolomite,limestone,mudstone,and sandstone.Experiments on rock fracture datasets in outcrop areas show that the improved model in this paper can accurately segment fractures in several rock backgrounds,such as dolomite,limestone,mudstone,and sandstone,and is more specific for edge segmentation of different types.The average pixel accuracy for rock fractures reaches 86.2%,and the average pixel accuracy for different rocks reaches 94.2%,respectively;The average intersection and merge ratio increased by 8.1% compared to the model before optimization.The experimental results show that the improved model in this paper has good segmentation performance in the field outcrop rock fracture segmentation task.