Research On Super-resolution Reconstruction Method Of Seismic Profile Image Combined With Attention Mechanism

Seismic profile images can visualize seismic data and show the formation structure in two dimensions(length and depth directions).Higher-quality images are conducive to the subsequent smooth interpretation of seismic data,but they are extremely difficult during seismic data acquisition.It is easily affected by the external environment and the performance of the sensor itself,resulting in a low-quality seismic profile image.In practical applications,a large number of sensors are usually needed to improve the resolution of seismic profile images,resulting in increased exploration costs.Therefore,this article makes an in-depth study on super-resolution reconstruction technology,and uses super-resolution reconstruction technology to directly reconstruct low-resolution seismic profile images into high-resolution seismic profile images,which is convenient and fast,while reducing exploration costs.Seismic profile images have obvious detailed texture features,which are often included in the high-frequency information of the image,and the existing super-resolution reconstruction network has the same learning ability for the low-frequency information and high-frequency information of the image,resulting in the quality of the reconstructed seismic profile image not tall.To this end,this paper proposes a super-resolution reconstruction method of seismic profile images combined with attention mechanism,the specific contents are as follows:(1)Aiming at the problem that the traditional super-resolution reconstruction network is not capable of learning high-frequency information of images,this paper proposes a multi-scale attention structure.First,convolution kernels of different sizes are used to extract image features,and then the high-frequency information is included.More channels are assigned with greater weights to realize the differential learning of the feature map by the network,and improve the ability of the super-resolution reconstruction network to learn the high-frequency details of the image.(2)In order to solve the problem of insufficient image feature extraction in shallow networks,this paper designs a densely connected block cascade network to increase the depth of the network.At the same time,the densely connected structure can effectively solve the problem that the deep network training is difficult to converge,and encourage feature reuse.Further improve the quality of reconstructed images.(3)In order to avoid the adverse effects on the network training due to improper learning rate setting,this paper uses Adaptive Moment Estimation as the network optimization algorithm,which can be adaptive during the network training Adjusting the learning rate is used to solve the contradiction between network loss and training time.The algorithm of this paper is implemented on the seismic profile data set,and several classic algorithms in the field of super-resolution reconstruction are used as a comparison.The experimental results show that the proposed algorithm reconstructs the image with clearer details and sharper edges,and also has objective evaluation indicators.It is superior to the comparison algorithm,which fully proves the effectiveness of the proposed algorithm.