Research On Semantic Segmentation Of Salt Body Image Based On Convolutional Neural Network

The accurate location of salt bodies is helpful to determine the location of oil and gas,which is of great importance to oil and gas exploration companies.The previous methods that relied on geophysical experts to interpret the images of salt bodies by hand were not only time-consuming but also subjective biased.With the rapid development of deep learning technology,convolutional neural network has excellent performance in acquiring image texture,color and other features by virtue of its superior network structure.In view of this,this thesis studies the semantic segmentation task of salt body image based on convolutional neural network.The main work of this thesis is as follows:(1)Aiming at the problems of small amount of salt image data,complex texture edge features and amorphous,a semantic segmentation model SEDense U-Net based on convolutional neural network is proposed by introducing attention mechanism and object context pooling strategy.On the basis of U-Net model,this model adds squeeze-excitation module and object context pooling mechanism to capture the correlation of the hidden state of encoder and decoder,and establish the relationship between two pixels in the long distance of salt body image.At the same time,jumping layer connection is added to combine shallow layer features with high-level information to strengthen feature reuse and transmission,and effectively alleviate the gradient disappearance problem.Experimental results on the salt body recognition challenge dataset published by TGS-NOPEC Geophysical Company show that the SEDense U-Net model is more accurate in the segmentation of salt body edge,texture and other features than the comparison model.Compared with the original U-Net model,SEDense U-Net model improves the average pixel accuracy by 1.17% and m Io U of the verification set by 1.58%.(2)In order to solve the problem of small amount of salt body data and possibly incorrect labeling,a new symmetric Lovász loss function was proposed by replacing the meaning of labels based on Lovász loss function.The function changes the input and output of the function on the basis of Lovász,and the null mask is treated as a non-null mask.The complete symmetric Lovász loss function includes the original Lovász loss and the symmetric Lovász loss.Finally,the output of the symmetric Lovász loss function is the result of adding the two parts and taking the average.Under the premise of not increasing the model complexity and the number of parameters,considering the true and false factors of predicted pixels,m Io U can be directly optimized in the neural network to measure the model performance more objectively.(3)According to the semantic segmentation task requirements of salt images and the requirements of model training convergence,a joint symmetric loss function was proposed by introducing binary cross entropy loss function.In the process of model training,the binary cross entropy loss function and the symmetric loss function were given different weights,and the two loss functions were propagated together to accelerate the convergence rate of the model and improve the model performance.Finally,Softmax function is used for classification and prediction.Experimental results on the salt body recognition challenge dataset show that,compared with the SEDense U-Net model using the cross entropy function,the average pixel accuracy of the SEDense U-Net model using the joint symmetric loss function on the verification set is improved by 0.54%,and the m Io U accuracy is improved by 0.62%.