Denoising research of petrographic thin section images with the global residual generative adversarial network

Petrographic thin section images have an important role in depositional environment inference, prediction of reservoir physical properties, and oil and gas analysis. To overcome the current challenges in thin section image denoising, we propose the global residual generative adversarial network (GR-GAN). Compared with the classical generative adversarial network (GAN), the residual network structure of the GR-GAN is reconstructed, and the loss function is redefined. The GR-GAN is then applied to denoise the thin section images in two different oilfields. The final denoising results confirmed that the GR-GAN achieves the best denoising effects on both visual evaluation metrics and objective evaluation metrics compared with colour block-matching 3D filtering (CBM3D), K-singular value decomposition (K-SVD), the GAN and a fast and flexible denoising network (FFDNet). Specifically, the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) generated by the GR-GAN on the test set are 28.2410 and 0.9674, 28.1075 and 0.9443, and 27.9919 and 0.9399, respectively, when the Gaussian noise is 15 dB, 25 dB and 35 dB, respectively, in the thin section image of the small-pore and fine-throat-type structures of J Oilfield; however, the data become 27.2841 and 0.9228, 26.8177 and 0.9162, and 26.3043 and 0.9068 for CBM3D, respectively, and these data generated by other methods are between the aforementioned two sets of data. The normalized root mean squared error (NRMSE) generated by the GR-GAN and CBM3D with the test set are 0.0327 and 0.1382, 0.0584 and 0.1341, and 0.0786 and 0.1382, respectively, when the Gaussian noise is 15 dB, 25 dB and 35 dB, respectively, and the NRMSE generated by the other methods is also between the aforementioned two sets of data. For other types of thin section images, when the Gaussian noise is 15 dB, 25 dB and 35 dB, respectively, CBM3D, K-SVD, the GAN, FFDNet and the GR-GAN show similar denoising effects as previously described. Moreover, in a denoising experiment repeated more than 10 times with the above methods, the GR-GAN has the shortest mean running time of 1.0589 s, and the mean running times of CBM3D, K-SVD, the GAN and FFDNet are 6.4609 s, 155.3158 s, 1.9394 s and 1.0622 s, respectively.