Digital Core Image Super Resolution Reconstruction Based On Deep Learning

In recent years,digital core is widely used in oil,mining and other fields due to data reusable,low cost,and other advantages.The digital core image resolution obtained by the CT equipment scan is inversely proportional to the core size.Therefore,the large-scale plunger sample images resolution obtained by the scan are low.It is impossible to meet the follow-up research demand of the core pores,cracks,rock skeletons,etc.To resolve this issue,this paper uses deep learning super-resolution reconstruction algorithms to reconstruct low-resolution digital core images to high-resolution images.First,a super-resolution reconstruction training data set is generated by performing image registration of the sub-sample digital core images and plunger sample digital core images.Because the cost of drilling core and the technical requirements are high,the sub-sample core data obtained in the laboratory is insufficient.This paper uses transfer learning method to successfully solve the problem of insufficient training sample data.Finally,the channel attention mechanism is added to the super-resolution reconstruction algorithm to achieve digital core image super resolution reconstruction.The specific research content is as follows:First,the digital core image feature extraction algorithm and FLANN match are combined to complete the rapid registration of the core images.This paper uses three algorithms to research image registration from different aspects with rotation,brightness,and translation.Through the results comparison analysis,the SIFT algorithm is more suitable for the digital core image registration when the brightness changes.The SURF algorithm is more suitable for the digital core image registration when translation or rotation changes.In any change,the registration result of the ORB algorithm is not as good as other algorithms.Therefore,SIFT and SURF algorithms are more suitable for the digital core image registration than the ORB algorithm.The two-dimensional space digital core image registration result and the high information of the sub-sample and the plunger-sample are combined to further complete threedimensional space digital core image registration.Second,because the sub-sample core data obtained in the laboratory is insufficient,the demand of training data quantity for deep learning super-resolution reconstruction cannot be reached.This paper uses plunger sample images and low-resolution images obtained by pooling of it to perform model pre-training.The training set is reconstructed on the basis of the pretraining model.The high-resolution sub-sample images and the corresponding low-resolution images in the plunger sample are training samples.Use the transfer learning method to retrain model,thereby effectively completing the super-resolution reconstruction of digital core images when training sample data is insufficient.Besides,in the super-resolution reconstruction,the channel attention mechanism is added to the SRRes Net algorithm to reconstruct high-resolution images that can clearly display information such as small pores,textures.