Research On Super-resolution Of Rock Image Based On Deep Learning

Micro-computed Tomography(Micro-CT)is an imaging technique that produces twodimensional images detailing the microstructure of porous rocks with which their physical properties can be studied.In order to obtain fine structure of rock samples,high resolution microscopic CT images of rock samples are often needed.The higher the resolution rate of an image,the more detailed information it contains,thus is more useful to the later image processing.The algorithm to improve the image resolution is also called image super-resolution reconstruction.This software-based method can improve the image resolution by using digital image processing technology without hardware upgrade.At present,the image super-resolution algorithm based on deep learning has been used widely because of their capability to solve the problems of the traditional algorithm,such as calculation complexity,blurred edges,loss of details and low generated image quality.In this paper,the deep learning based super resolution method is studied thoroughly,and two different image super-resolution algorithms,which are based on and not based the generative adversarial network framework respectively,have been proposed,with their application to rock microscopic data sets.The main contributions are as follows:(1)A super resolution algorithm for generating adversarial network based on asymmetric residual blocks is proposed.Asymmetric residuals block can better integrate the features generated by the shallow convolutional network,while the application of global residuals learning strategy can transfer the shallow features to the end of the network much better,the combination can make the feature fusion more sufficient as we expect.By adding asymmetric residual blocks to the SRGAN generator network and the optimization,the problem of insufficient SRGAN feature fusion is solved.Compared with SRGAN,the proposed algorithm can improve the two objective evaluation indexes of PSNR and SSIM to a certain extent,and the image texture is better restored.(2)A back-projection network super resolution algorithm based on feedback module(FBPSRN)is proposed.The algorithm is based on the Deep Back-projection Network(DBPN),and the feedback module is composed of the upper sampling block and the lower sampling block alternately.The high order features extracted from the deep network are applied to the learning of the low order features through feedback,which solves the problem of low parameters utilization efficiency of the DBPN network.Since generative adversarial network is not used with the algorithm,artifacts induced by SRGAN can be avoided.Experimental results on rock microscopic data sets show that,compared with DBPN,the image details recovered by FBPSRN are optimized,and both PSNR and SSIM indexes are improved.In summary,super-resolution algorithm based on generative adversarial network has a better texture restoration quality while the PSNR and SSIM performances are relatively poor compare to that of none-GAN based method.The none-GAN based method has the advantages of artifacts removing and image details recovering.