<p>Focused Ion Beam Scanning Electron Microscopy (FIB-SEM) technology is a crucial method for obtaining three-dimensional (3D) nanoscale digital rock images. However, due to surface topography or compositional differences in rock samples, “curtain noise” appears in FIB-SEM images, degrading their quality. Traditional image filtering and denoising methods can improve FIB-SEM image quality but are ineffective at fully removing “curtain noise”. To address the issue, this paper proposes two Convolutional Neural Networks (CNNs) for curtain noise denoising: U-Net: Convolutional Networks for Biomedical Image Segmentation (U-Net) and Denoising Convolutional Neural Network (DnCNN). Experiments evaluate the denoising performance of these models using digital rock images obtained from FIB-SEM equipment. The results show that both models significantly enhance image quality. Based on the updated values, the Peak Signal-to-Noise Ratio (PSNR) for the U-Net and DnCNN models are 29.10dB and 27.28dB, respectively. The Structural Similarity Index Measure (SSIM) values are 0.73 and 0.38, and the Learned Perceptual Image Patch Similarity (LPIPS) values are 0.31 and 0.45. After denoising with both models, the estimated porosity of the digital rock images is closer to the true value. For a rock image with a manually segmented porosity of 6.3%, the threshold-segmented porosity decreases from 7.9% to 6.1% and 6.0% following curtain noise removal by U-Net and DnCNN network, respectively. The U-Net model demonstrates strong denoising performance while preserving texture and details, whereas the DnCNN model significantly reduces the curtain effect but introduces noticeable smoothing and blurring. Moreover, the generalization ability of the DnCNN model is relatively weaker than that of the U-Net, making it less effective when applied to unseen FIB-SEM images. This research provides a foundation for the accurate reconstruction of 3D nanoscale digital rock models.</p>