Targeting the limitations of existing image denoising methods, GAN-DFFE (generative adversarial network with dual-frequency feature extraction) was proposed. In the GAN-DFFE, the generator employs octave convolution to transform local redundant channels and extract dual-frequency features from images. The channel attention residual module was utilized to capture key features. The discriminator leverages a Markov model, while the final network layer adopts Leaky ReLU as the activation function to address the issue of dead neurons. The GAN-DFFE was applied to the denoising of petrographic thin section images from J Oilfield in Jiangsu Province and S Oilfield in Shaanxi Province. Compared with other denoising methods such as EPLL (expected patch log likelihood), CBDNet (convolutional blind denoising network), and IRCNN (inception-resNet convolutional neural network), it demonstrates superior performance. Specifically, for standard deviation levels of Gaussian noise of 15 dB, 25 dB, and 35 dB, the average PSNR, NRMSE, and SSIM values obtained using the GAN-DFFE on thin section images of small-hole and fine-throat from J Oilfield are 30.140 0 dB, 0.031 9, and 0.973 9, respectively, which outperform those of EPLL, CBDNet, and IRCNN. For other types of thin section images, the GAN-DFFE also exhibits better denoising effects. This method can effectively remove noise from thin section images and provide robust technical support for geological researchers.